19.09.2023 - 16.01.2024

  • Added:
  • ○ New worker - Device updater, used to automatically update the
  • rented devices
  • ○ Information about the current version, the latest available and the minimum
  • supported in the "Update" tab on the personal device
  • ○ Display nickname in the Settings view
  • ○ Requirement of minimum acceptable number of publication images
  • ○ Added "processing" field instead of "0" in "Conversion list" during
  • conversion processing
  • ○ Unification of Back button
  • ○ Conversion status information
  • ○ Redirection from training notifications to model list
  • ○ Added author in dataset details on Shared and Marketplace views
  • ○ Enhanced filtering of models and templates (framework/pretrained model)
  • ○ Automatic search after changing filter in Unsplash upload
  • ○ Checking if the number of images in the dataset is suitable for publishing before
  • the Publish modal
  • ○ Introduced skeletons
  • ○ Adding the ability to open datasheets, models, etc. in a new window using Ctrl +
  • LPM and the middle mouse button
  • ○ Unit tests for existing functionalities
  • ○ Fully functional breadcrumbs, which are currently the basis of navigation
  • ○ Added Select with the ability to add categories
  • ○ Added cloning tables to reports
  • ○ Added ability to remove Dataset from details view
  • ○ Ability to combine multiple categories into one
  • Updated:

○ Improved validation of sent messages in the "Help" tab

○ Redesigned and improved category linking system

○ Improved access token view for lower resolution than FullHD

○ Changed indexing throughout the frontend application - now all

pages start with "1" instead of "0"

○ Removed Back button from all non-wizard sites

○ Improved performance and rendering speed of Angular 16

○ Stretching small images in marketplace gallery

○ Improved the layout of the modal that allows AIDWS updates

○ In template/models, removed unnecessary frameworks in filters

○ Routing refactor

○ Improved CSV file generation

○ Fixed sidebar shadow (was glowing in dark mode)

○ Added in the cogwheel the option to edit the Dataset name

○ Errors related to not closing some frontend components

○ Fixed and edited performance of Drag and drop function in the Upload view

○ Added model type information in training details view

○ Graphic fixes in Marketplace view

○ Added existing categories in Dataset during File Upload

○ Fixed missing steps and placed Tutorial Mode steps

○ Added restriction of Subdivisions parameter in Detection training parameterization

  • Fixed:

○ Fixes when deleting multiple datasets

○ Various kinds of errors in uploading files to S3 and downloading them

○ Improvements of Dashboard view, Dataset details, Templates, Models, Image

editor and Manipulator on mobile devices

○ Removed duplicate fields in pl.json

○ When first entering Marketplace or Shared, the first page is already

marked in the pagination

○ Selecting a tab that has a subtab correctly selects the tab in the sidebar

○ Date selection fields in Reports view

○ Send button on Help view (disabled and loading status)

○ Bottom panel with Back button covered the items below it

○ Category tab hovered over buttons on Dataset details view

○ Responsiveness of forms used in Templates and Models view

○ Unlocking of dataset after manipulation ending in error

○ Fixed pagination in live_testing and training/template

○ Window for typing messages in help can no longer be enlarged infinitely

○ Pending datasets can no longer be accessed

○ Fixed strange behavior in resize in dataset manipulation

○ Displaying wrong values of not uploaded images to dataset

○ After returning from image-editor, gallery in dataset no longer stretches "under"


○ Fixed incorrect addresses for CRUD operations in category editing

○ No translations in training parameterization

○ Correction of drop-down Menu when selecting categories

○ Fixed disconnected WebSocket connection

○ Set rights by author of shared/publicated datasets are now strictly respected

○ Graphic fixes in Unsplash uploading

○ Graphic fixes in training detail view

○ Fixed locked "Edit" button when changing categories for multiple images

○ Unification of the reset button for Inputs in the My Notifications tab

○ Fixed a bug related to deleting categories after entering Image Editor

○ Added missing Skeletons and Load Indicators in missing locations

○ Removed redundant Requests in Marketplace and Shared datastes view

○ Blocked the ability to add and edit categories in Datasets shared to a user without

the right to modify

○ Supporting gallery exit in Unsplash Upload view by adding a "Escape" keyboard

shortcut and added a button in the upper-right corner of the screen

○ Fixed incorrect captions in the Help tab

○ Disjointed Inputs and Labels in the Manipulation Operations view

○ No return to previous page when removing devices from Live Testing

○ Empty account - clicking "Go to datasets" no longer redirects to an empty page



19.09.2023 - 16.01.2024

  • Added:

○ Implemented versioning system and support for related edge cases

○ Confirmation of file deletion on Results view

○ Enhanced model filtering (framework/pretrained model)

○ Support in AICS for multiple sessions

○ Support in AICS for multiple WebSocket clients listening to a single session in the

context of video streams and status

○ Enhanced security of use through shorter JWT token times and implementation of

token refresh mechanism

○ Support for Pytorch models on Google Coral

○ Enhanced system for uploading files to OSAI from streams

○ Tooltips to explain inaccuracies in frame upload to OSAI

○ Added scrolling on mobile devices in player

○ Added 'Back' button when selecting a model when there are more than 1


○ Added a button to Upload files in the Saved Inputs view.

○ Supported mechanism for refreshing authorization tokens during use of the


○ Downloading missing model from OSAI when entering from "Run model" tab

instead of error information

○ Information on whether the model was re-trained

○ Updating the model name after it was changed in OSAI

  • Updated:

○ Redesigned AICS to provide greater stability and performance

○ Improved authorization mechanisms at system startup

○ Improved real-time communication between AICS - Backend and AICS - Frontend

○ Eliminated multiple sources of memory leaks

○ Improved error handling during ongoing streaming

○ Unification of the Back button

○ Updating the player for the new AICS

○ Improved performance and rendering speed of Angular 16

○ Fixed and standardized gallery performance in AIDWS in input/output

○ Handled an error when trying to run a model on a currently occupied stream, by

displaying a modal window with an appropriate informational message

○ Added default values for Leopard camera parameterization.

○ Improved display of various types of errors

  • Fixed:

○ Application freezes after file upload error

○ Fixes in the player for displaying incorrect time and filling up of the buffer

○ Uploading always the same Interval value

○ Loading dark mode

○ Back button panel covered the items below

○ Photo/video deletion notification scrolled with the screen

○ Graphic improvements to buttons in camera view

○ Graphic improvements to Wizards

○ Improved error handling related to temporary camera disconnection while live

streaming is in progress

○ Graphical unification of the sidebar - color and shading fixes

○ Elimination of dependencies between Inputs and Outputs causing various types of


○ Error with handling the case when authorization in OSAI fails

○ Errors related to uploading large files in PNG format and 4K resolution



19.06.2023 - 19.09.2023

  • Added:

○ New model tiles in the Predictor wizard.

○ Added marketplace with a system for viewing and buying datasets. This includes a detailed view of datasets for purchase with a new gallery, descriptions, and a modified category management system. The list of public datasets includes filtering and sorting options: by price, number of purchases, number of downloads, and number of images.

○ Added the ability to make datasets public.

○ Added a view of the publication list for a given dataset.

○ Added the ability to share datasets.

○ Created a view of datasets shared with the current user, with the option to accept or decline pending shares.

○ Added a view of all shares of a given dataset.

○ Added ability to clone datasets along with support for webhooks.

○ Added field with information about image extension in image details.

○ Breakdown of the side navigation in the Datasets tab into sub-tabs: Owned, Marketplace and Shared.

○ Restricted access to dataset options depending on the assigned perimissons in shared datasets, three levels of permissions have been introduced: clone ability, edit ability or read only.

○ Introduced nicknames for users to identify themselves in the marketplace.

○ Implementation of a subsystem that checks the version of AIDWS on devices and determines if updates are available/required for it.

○ New implementation of a predictor based on Tensorflow Object Detection.

○ Added a message to warn that going back from the merge view to the classification/detection selection will delete the entered data.

○ Ability to retrieve dataset size.

○ System for merging two categories into one from the dataset detail view.

  • Updated:

○ Deleting images in the dataset detail view has been improved. Currently, when deleting, the dataset is not completely reloaded, only the missing images are reloaded.

○ Modified gallery in Unsplash view.

○ Improved data security with additional protections at the level of images, datasets, categories, labels, etc.

○ Increased notifications informing about key changes in the system requested by the user. ○ Renamed 'epochs' to 'batches' in Templates and Models in Detection/Tensorflow.

○ Improved the appearance of the 'Show/hide' button in Bilings view.

○ Fixed formatting of too long notifications.

○ Translation fixes.

○ Showing different information in dataset details depending on the view.

○ Live Testing tab URL changed from "deployment" to "live_testing".

○ Image extension information in image details.

○ Eliminated memory leaks that prevented large datasets from being uploaded and destabilized the backend.

○ Improved error handling in queuing systems.

○ Optimized system memory usage.

  • Fixed:

○ Fixed incorrect redirection of the Back button.

○ Fixed incorrect notification after deleting a rental device.

○ When creating and editing a new template, we can no longer access the parameters without entering the template name.

○ Fixed camera access when access is denied.

○ API fixes related to requests for non-existent objects.

○ Fixed a bug that caused the Polish language to be incorrectly initialized when logging in.

○ Fixed clipping of the photo when doing manipulations in the manipulator.

○ Preservation of photo proportions in manipulator operation views.

○ Search engine in My datasets now searches correctly and does not redirect to the dashboard when filtering is applied.

○ Fixed misalignment of tables on the dashboard.

○ Deleting a label no longer disables keyboard shortcuts in the image editor.

○ Deleting a photo in the image editor no longer causes errors.

○ Fixed the Back button in Datasets in the Polish version, which redirected to the wrong view.

○ Fixed a progress bar in the 'live_testing' that could load backwards.

○ Predictor and manipulator work when entering datasets via notifications.

○ Predictor supports outdated models - from now on it sends csv files to S3 as in newer models.

○ Selecting the first image in dataset details no longer causes backward pagination to fail.

○ Images in model training details no longer load indefinitely when given images are removed from datasets.

○ Added load indicators in the model and template lists view when paging.

○ Handling of validation errors on the frontend side.

○ Fixed bug with blocking dataset editing after starting training.

○ Fixed bug with drawing labels on another image when switching quickly in the image editor view.



19.06.2023 - 19.09.2023

  • Added:

○ Added a new tile component in the models view.

○ Added saving of the current page in the models view.

○ Redesigned the live-testing view into a wizard form with individual steps outlined.

○ Validators for all requests.

○ Added loading spinner in 'Cameras' view.

○ Added filtering in the list of models.

○ Corrected error handling for unloaded subsystems and lack of authorization from OSAI.

  • Updated:

○ Significantly improved stability and performance of the backend and stream processing service.

○ Better communication with the frontend and faster response time to streaming interruptions on the backend side.

○ Better handling of streaming errors.

○ Better handling of other types of exceptions from the backend.

○ Completely rebuild frontend-backend communication in terms of all requests for models and running them on photo/video/camera.

○ Fixes for warnings about exiting the view while uploading from a camera without a model is in progress.

○ Improved login security.

○ In case of a non-existent device, the system redirects to the login page.

○ Correct display of outdated system version.

○ The action button can no longer be clicked when the camera is streaming.

○ Blocked access to login view when the user is logged in.

○ Improved handling of cameras with internal errors.

  • Fixed:

○ Eliminated errors causing stream processing freezes.

○ Eliminated errors causing critical stream processing errors.

○ Error causing that after deleting an input, the output based on it could not be previewed.

○ Eliminated errors related to downloading models.

○ Eliminated errors related to uploading images to OSAI.

○ Improved recording of frames every time interval - better synchronization.

○ Running the model on the photo and running the camera preview.

○ Upload of dataset from video was not responding to changed configuration.

○ The 'Stop stream' warning is displayed correctly only when the stream is not fully loaded, and not every time when exiting the player.

○ Logging out and logging back in no longer disables dark mode.

○ The 'Settings' button responds every time it is clicked.

○ Fixed erroneous redirection to the login page right after token authorization.

○ Pressing the logout button correctly invokes a modal that warns that streaming has ended. ○ Translation fixes.

○ Errors related to displaying inputs and results.

○ Exception handling for the same dataset name when uploading images on OSAI.

○ Sorting of detection/classification results is done by their creation date.

○ First login correctly transfers the color style of the page from OSAI.



28.03.2023 - 19.06.2023

  • Added: 

○ Portal translation into Polish

○ New authorization system on the portal, which eliminates previous problems, simplifies work with OSAI using the API, and does not require constant user login (everything is done in the background)

○ New system for verification of external systems with additional requirements regarding their minimum version

○ Implemented webhook system to improve working with OSAI using the API

○ Added the ability to preview dataset size

○ Added basic information in the detailed dataset view: size, numer of images, creation date

○ Added basic information in the detailed models view: configuration, creation date

○ Added API tab in the sidebar 

○ Added support for RMSprop, Adadelta, Adagrad, Adamax, Nadam, Ftrl optimizers in Keras training

○ Covering a significant part of the code with appropriate tests 

  • Updated:

○ Improved worker performance

○ Significantly increased security level on the portal

○ Changed default number of epochs in trainings

○ Redesigned dataset details view – drop-down category filter bar on the right-hand side of the screen 

○ Expanded model information when downloading with additional fields

○ Changed the order in the training wizard: first select datasets and categories, then view category merging and configuration

○ Rebuilt the logic for the tutorial mode, moving files to directives, separating tutorial logic from components

○ Filtering of the pretrained models in the models and templates view

○ Rebuilt logic for automatic training, optimization, not sending unnecessary fields

○ Rebuilt validators for epoch size in training and min and max in Tensorflow

○ Added creation time in Access Token view

○ Updated device connection instructions 

○ Changed the order in the wizard for adding a new Device - first select the device, then enter the name and URL

○ Added tooltips for Data Augmentation in Tensorflow 2.11

○ Adjusted tile views for the Polish language version

○ Changed the default pretrained model for the Darknet framework in automatic training to Yolo V4 Tiny

○ Changed icon colors in the Image editor

○ Improved shades of black in Dark Mode

○ Disabled hover animations for unsupported conversion machines

  • Fixed:

○ Errors when merging categories due to categories being saved and incorrectly assigned

○ Error causing the maximum download of 100 photos from Unsplash

○ Errors resulting from repeatedly received statuses from workers

○ Errors in registration using a Microsoft account

○ Saving filters in models and templates when refreshing the page

○ Fixed incorrect counting of the maximum number of epochs

○ Fixed the progress circle in the tile view, unifying and displaying in the middle of image 

○ Fixed tutorial mode – in case of a non-existent selector in which the help window would eventually be displayed

○ Incorrect batch size value in templates

○ No backgorund for filter drop-down button in the dataset

○ Fixed blocking flow when adding and editing templates 

○ Fixed training wizard – when navigating Back, Next, the old values remained in the configuration form

○ No redirect from wrong URL on dashboard

○ Improved field names in PyTorch augmentation

○ Fixed error when passing null instead of categories ready for training – training configuration 

○ Incorrect max batches set when editing Templates

○ Incorrect loading of values in template after refresh

○ Incorrect image limit display in Unsplash and unresponsive filters when editing and deleting

○ No saving of color options

○ Improved status display during conversion

○ Updating the number of images and other dataset parameters after each image deletion          

○ Errors when converting to Google Coral

○ Errors with image handling in the dataset manipulator



28.03.2023 - 19.06.2023

  • Added:

○ New authorization system compatible with OSAI, appropriate reminders in case of outdated system version, system self-authorizes on OSAI

  • Updated:

○ Rebuilt model download to be more stable and responsive

○ Rebuilt HTTP connections - improved stability

○ Improved status handling during live video preview

○ Updated internal libraries

  • Fixed:

○ Login errors



25.01.2023 - 28.03.2023

  • Added: 

○ Integration with Unsplash as a third-party image provider for datasets

○ Support for the Tensorflow Object Detection API

○ Filtering, sorting and pagination in the Live Testing view

○ Pagination in the device selection view at model startup

○ Counting category usage in detection and classification contexts

○ Views under Tensorflow detection training configuration 

○ New category selection view for training, taking into account categories that cannot be used in the context of a given training type, as well as signaling categories that do not meet the requirements defined by Unsplash 

○ Option to resize thumbnail images in the dataset detail view

○ Views for creating a dataset and uploading images from the Unsplash platform

○ Precise labeling of dataset types (Personal and Unsplash)

○ Ability to preview image details from the dataset details view

○ Displaying information on the model tile about whether it is re-trained and the name of its parent

○ Tooltips in the tiled view of models showing all datasets and categories used in training

○ Filter models by whether they are re-trained or not

○ Option to list the sizes of all conversions for a model in the gear wheel of its tile 

○ Storing filters in the training view

  • Updated:

○  Rebuilt algorithm for separating images into training, validation and test subsets depending on the split percentage - eliminated edge cases causing underestimation of some categories and guaranteed better split for detection categories

○ Saving the last model with the correct accuracy if it reaches 0% in the Darknet framework

○ Support for uploading image archives with duplicate names and correct classification categories taking into account the directory structure

○ Changed the return icon in the labeling view

○ Login security increased

○ Hidden tokens in Live Testing, which are now only available on request until a connection is made to the device

○ Start charging rental time to the device when it actually becomes ready, not when it is created in the app

○ Disabling unnecessary service causing unnecessary load on the system

○ Optimized updating and removal of tasks running in the background

○ Rebuilt and optimized archive upload module

○ Increased system stability by covering more of the code with tests

○ More accurate information when trying to upload to a non-existent dataset

○ Added a new Checkbox type to the shared component and used it in the Dataset selection view during training and in the dataset detail view during category filtering

○ Optimized generation of uploaded images error report to CSV file

○ Ability to preview errors in the Recent Errors table

○ Ability to enable re-train from training statistics view

○ Added filter to dataset view - Personal / Unsplash

○ Moved Tooltip with model configuration in Tile view

○ Improved stability and additionally secured connections to worker queues

○ Improved ‘Resize’ layout in the manipulator view,

  • Fixed:

○ Errors when displaying accuracy/loss charts

○ Errors in API method descriptions

○ Errors related to merging categories for re-training

○ Error preventing removal of Access Token

○ Errors when uploading some images related to their extensions

○ Error when trying to re-train a model where datasets have been removed

○ Error when using the Adam optimizer in Tensorflow Keras

○ Error related to the creation of Access Token with the same name (uniqueness at the user level, not the entire portal)

○ Incomplete device signature when renting a device - Live testing view

○ Error with keyboard shortcuts in the new labeling view - accidental activation of ‘Invert axes’ option

○ Redirection from ‘Recent predictions’ table to detail view of completed prediction

○ Code refactoring when searching existing elements in the table on the Frontend side - avoiding errors in the console when referencing fields for non-existent objects

○ No warning before creating a Dataset with the same name

○ Deleting all reports on a given page resulted in inability to use pagination

○ Going to conversion from notification or task triggered an error

○ Errors when adding new predictions related to repeating categories

○ When resizing thumbnails in dataset detail view, drawn labels were blurred

○ Names of frameworks and models are now mapped by name so they will always have the same background color

○ Input when creating a new dataset was losing focus

○ In the Advanced section when parameterizing training, if you went back to the previous Wizard page, previously set values were replaced with default values from the Pretrained Model configuration

○ Updated documentation addresses in the Training view

○ Stylistic improvements to the Help window

○ Set the minimum number of epochs for Tensorflow to 1000

○ When entering the ‘Model History’ tab on the ‘Model Details’ page, the model history was not displayed

○ Color corrections in dark mode tiles

○ Stylistic corrections of Tags - uniform background color for all tiles



25.01.2023 - 28.03.2023

  • Added: 

○ Support for models with TensorFlow Object Detection API.

  • Updated:

○ Performance optimization for models with TensorFlow/Keras

  • Fixed:

○ Errors when loading some large models on a stream or video file


15.09.2022 - 25.01.2023

  • Added: 

○ Public API for the backend with detailed descriptions of the endpoints and their parameters

○ A system for automatic and detailed reporting of errors directly to developers covering all elements of the OSAI portal, allowing errors to be caught and fixed much more quickly and improving user-support communication

○ Deletion of multiple datasets at once o New endpoints for uploading photos

○ A system for checking data inconsistencies between the database and the S3 storage - if missing files are detected on S3, they are also deleted from the database with the date of the previous check as the deletion date, so that the user is not charged for this

○ Possibility to change the model name

○ Support for ZIP archives in non-standard encodings

○ Deletion of a photo from the image editor

○ New layout added to login view
○ New labeling mode - moving the side menu to the right side in the form of a drop-down panel. It is also possible to switch between two modes of adding new labels - classic mode with a drop-down window for selecting a name when creating a new label, and automatic mode, which allows automatic creation of annotations with the initially declared category name

○ New look and animation of tiles throughout the application (Datasets, Models, Templates, Devices)
○ Conversion view in model details redesigned as a tiled view o “My Devices” and “Add New Device” views redesigned as a tiled view o Upload error list added to notification view

  • Updated:

○ Upgrade from .NET 6 to .NET 7

○ Significant performance improvements, optimized queries to API and database. Faster updating and deleting of data

○ More accurate backend reporting of errors encountered

○ Background data download in the image editor to reduce reload times for subsequent images

○ Caching of uploaded images in the dataset detail view, making browsing much faster and avoiding having to download the same elements again

○ Redesigned upload of images from archives. New algorithms have been introduced to better pair images with their annotations for all supported formats, taking into account their classification categories. These categories are fixed if they contain forbidden characters

○ Modified JWT Token for authentication

○ Significantly improved validation of data entered into the backend

○ Added spinners when loading data in notifications

○ Redesigned reporting of incorrect upload images - added paging, protection against downloading too much data at once, export to .CSV file is now on the backend

○ Redesigned notifications - details translation is now on the frontend, providing flexibility to add translations in other languages
○ Limited size of uploaded images

○ Added archive file size limit when uploading images

○ Added option in edit mode to remove a classification category

○ Changed color of tiles in "Dark mode"

○ Introduced a mode suggesting the unavailability of a device when deleting it

○ Validation of fields in the Add new Device view

○ Styling of 'help' type icons in Add new training view

○ Changed field names and adapted to the configuration available on the server side in the Pytorch training form

○ Added Debounce Time when filtering by name in the models view to avoid too frequent queries to the Backend

○ Introduced an option to add a new category in the 'Category Manager' view in Dataset Details

○ Confirmation of notification deletion

  • Fixed:

○ Incorrect display of reports during generation

○ Errors when displaying model history

○ Errors related to clearing resources on the backend

○ Errors related to device rental

○ Error causing the system to freeze when a lot of images with errors were included in the notification

○ Error related to ColorJitter in training configuration

○ Error related to RandomSolarize in training configuration

○ Error causing predictions not to be included in automatic reports o Incorrect sorting of reports

○ Removed unnecessary tutorial point in "Connect new device" view

○ Missing spinner when loading dataset page

○ Wrapping the device name so that it does not go beyond the tile

○ It was possible to press the "delete" option while the device was being released

○ Incorrect merging of classes when adding a new training

○ Incorrect saving of Templates - fixed incorrect saving of fields and then correct completion of forms

○ Fixed validation errors when creating a new training

○ Error with ability to delete datasets being processed

○ Improved appearance of report table

○ Error with pagination when deleting notifications

○ When uploading photos in Live Camera mode is interrupted, it is not possible to save them

○ Adjustment of the view to mobile mode



17.09.2022 - 25.01.2023

  • Added: 

○ Parameterization of the maximum number of sessions processed simultaneously in AICS

  • Updated:

○ Upgrade from .NET 6 to .NET 7

○ Integration of automatic and detailed error reporting into the system

○ Adaptation to new OSAI API

○ Improved OSAI error handling

○ Redesigned image upload to OSAI, improved start and end mechanism for uploads

○ Introduced automatic selection when one device is available

  • Fixed:

○ Eliminated freezing moments when uploading photos to OSAI

○Locked action buttons in the view of available cameras while they are in use

○ Back button during image upload was causing the upload to be canceled (ultimately confirm leaving the view)

○ Ability to switch between pages if the password has not been changed

○ Locking the Upload window during "Don't upload" mode.



25.07.2022 - 15.09.2022

  • Added:

○ Added a queue item in the training tile view for pending tasks

○ When adding a category during file upload, completing the category field in Object tracker view

○ Added view with Access Token generation

○ The "Settings" tab has been rebuilt

○ Completely new system of lending and releasing devices with corresponding Tasks running in the background and their progress expressed in percentages

○ Pagination, sorting and filtering of notifications

○ Filtering and sorting of devices in the Live Testing tab

○ Filtering and sorting of reports

○ Support for uploading archives in .tar and .tar.gz

○ Added additional points on the graph drawing from the epoch whose result was rejected (the best model was trained in an earlier one)

○ Added validation per name in the category management view

○ Use of new tables (changed filtering, sorting, show/hide columns) in notifications, reports and live-testing

  • Updated:

○ Blocked in the object tracker the options to rotate and change the shape of the marking

○ Completed training parameterization of the 'batch size' field with validation, i.e. the 'subdivision' field cannot be larger than the 'batch size' field

○ Added scroll to "Labels" section in Image Editor view

○ Significant performance gain for Dataset Manipulator

○ Saving the history of the entire training (including rejected epochs)

○ Sorting by category name in datasets and labeling view

○ Performance improvements throughout the system

○ Visual improvements

○ Changed instructions in my devices

○ Changed delay on filtering to one second in models view

○ Adjusted disabled on the delete button for models

○ Changed navigation for the back button in deleting images in datasets

○ Changed the logic of the authorization mechanism in the system

○ Validation when adding a new token in settings view

  • Fixed:

○ Updated progress bar for conversions when shuffle is finished in models view,

○ Template name missing when editing,

○ Badly displayed upload bar in "Tasks" tab. - Improvement for 99 and 100 percent values,

○ Added description of "Crop threshold" field,

○ Moved to appropriate view when clicking on a finished conversion task

○ Errors causing incorrect object annotations

○ Errors causing manipulation to hang

○ JPEG codec errors causing manipulation to abort

○ Errors causing datasets and models to remain in use even though the tasks associated with them were completed

○ Errors related to authentication (login and logout) and cookie security,

○ Bug causing object annotations to be saved incorrectly when zooming in and out of photos

○ Disappearing auxiliary lines in labeling view

○ Errors during dataset download related to empty annotation files

○ Errors when uploading ZIPs

○ Errors during conversion

○ Memory leaks

○ Re-train not showing old datasets



25.07.2022 - 17.09.2022

  • Added:

○ Rebuilt the Live Testing view into a wizard form, with tiles representing the

○ Hardware and the current configuration visible on the right side

○ Added loading icons at data loading moments to improve the responsiveness of the page

○ Added validation when changing passwords to the same

○ Added support for Leopard Imaging camera

○ Added detailed information in tiled view of models

○ Added support for ArduCam with IMX477 sensor hooked up to CSI → USB adapter

○ Pagination, sorting and filtering of models

○ Supported firing of Pytorch models from Movidius

○ Protection against multiple firing of streams

○ Rebuilt results view, extracting logic to separate components, applying new styles to tabs on the view

○ Added Roboto font throughout the application, customized font in buttons

○ Added result-card component, which is reused throughout the application as a tile with photo, model name, etc.

○ Pagination in the result(input/output) view

  • Updated:

○ Displayed more information in the "Results" tab (model, hardware and creation date),

○ Displayed classes that recognize the model in the tiled view,

○ Rebuilt stream file upload to be more stable and responsive,

○ Width of drawn bounding boxes in detection network results is now dependent on source resolution

○ Reduced latency in the image displayed in the camera preview

○ Improved stability and system performance

○ Improved reading of the resolution of the photo/video uploaded to the portal

○ Rebuilt communication between subsystems during streaming

○ Handling of errors returned during communication with OSAI

○ Improved closing of inactive streams

○ Improved responsiveness of views,

○ Changed popover in tile details in live-testing view

○ Changed the logic of the authorization mechanism in the system

○ Redirection after refreshing stream view to input view

○ Adjustment of responsiveness in images-gallery view

  • Fixed:

○ Removal of unnecessary pop-ups for authorization errors,

○ Correct validation for the same dataset name when uploading camera images,

○ Ability to preview video when uploading custom video file,

○ Thumbnail loading in tiled view "Results",

○ Correct validation in the absence of the selected hardware and model,

○ Navigation to main views in case of missing parameters in navigation,

○ Correct handling of websocket connection after correct login,

○ Improved width and height validation when uploading images from Live stream,

○ Improved tiled views after page refresh (no model details returned),

○ Support for uploading a photo when dimensions are too large,

○ Artifacts resulting from drawing bounding boxes on objects after model processing in video

○ Validation of labels during file uploding

○ Improved uploading of files

○ Fixed showing the number of photos uploaded on OSAI

○ Fixed bug with disappearing hardware when changing theme

○ Fixed bug with deletion of downloaded models during page scrolling

○ Incorrect scaling on camera preview view

○ Name and percentages in low-resolution photos in result-card

○ My cameras - the list was too wide (scroll appeared)

○ Without having a camera we could go to url /live-testing/live-view

○ The name of the new camera could have one character

○ Deleting the last item in results didn't show the information about missing data

○ Fix next/previous in image gallery preview

○ Bug with redundant toastr when logging in/out

○ Error when returning to input-selection

○ No response when clicking on a tile on hardware selection



22.07.2022 - 25.07.2022

  • Fixed:

○ Calling the camera in Object Tracker did not work in the Mozilla Firefox browser

○ When switching between tabs in the model's detail view, tutorial windows could duplicate and cause shifts on the page



08.07.2022 - 22.07.2022

  • Added:

○ Support for uploading annotation files without extensions (Pascal VOC without .xml and Darknet TXT without .txt along with the labels file)

○ Labeling has an added axis reversal mode for moving the view (with the keyboard shortcut "I") and a mode where you always create a new label, even if you are in an area where there is already another one (with the keyboard shortcut "W")

○ Notifications and appropriate information in the lower left corner of the portal when there is no Internet connection and when it is restored

○ Parameterization of the docker-compose.yaml file for different production servers

  • Updated:

○ Improved performance of models trained in the Pytorch framework

○ Unified flow for all workers on the portal (training, prediction, conversion and manipulation)

○ Improved checking of entered configuration parameters before training and post-training in all frameworks

○ Predictor wizard immediately counts images that would be subjected to prediction in the currently selected mode and displays this to the user

○ Impose locks on datasets used in training and on datasets and models used in the predictor, making it impossible to edit or delete them

○ Imposition of labeling mode locks for datasets used, making it impossible to edit the labels, only to read them

○ Keyboard shortcuts are already there for all possible operations in labeling mode

○ Editing labels by dragging them by their vertices now defaults to the mode without preserving proportions. Preserving aspect ratio is possible by pressing the "Shift" key.

○ Logging out on one tab redirects to the login window on all tabs open within that session

Improved flow for internal workers and improved error logging

○ Additional security for internal workers to prevent RAM overrun errors,

○ Re-downloading of data in tables on dashboard in case of restoring internet connection

○ Adjustment of category and config display in models and templates view (json file formatting and window height/width adjustment)

○ In the case of too long name in the tiled view of Dataset and Templats, collapsing the text and adding a Tooltip with the full name, which is activated when the mouse cursor hovers over the name

○ Adding a reminder window for necessary camera access permission for Upload from LIVE broadcast

○ Changed keyboard shortcuts view in labeling view

  • Fixed:

○ Image extensions ending in uppercase letters caused errors in internal systems

○ Predictor with Pytorch model was not detecting categories

○ Failing to scale images and annotations that do not meet portal requirements (below 8x8 px or above 4096x3112 px)

○ In detection training, selecting one label category from a dataset resulted in deleting all others belonging to that dataset from the database.

○ The device rental algorithm has been improved. Previously, it could randomly rent a device with bad hardware

○ Correct selection of all categories in the category manager view, using the "Check all" checkbox

○ When uploading a dataset as a ZIP, the record with progress showed up in the Recently used datasets table

○ No dataset category after starting to upload files to the dataset

○ Changing the photo category in labeling mode in the dataset did not affect the progress in the upper right corner

○ When preparing a rent device, the icon with the animation overlapped the text when the screen resolution was reduced

○ Tutorial on the dashboard was reappearing when updating any of the processes.

○ Fixed drawing of labels in Upload view, previously when importing photos from LIVE broadcast several times, labels were not visible

○ When selecting a template when creating a new model, it removed the previously specified name - overwrote it with the template name

○ Blocked deletion of values in selecting a framework for training - this field must always be filled in

○ Redirection from shuffles and notifications - fixed for completed uploads and predictions

○ Fixed deletion of imported images from LIVE camera

○ Correction of single label in tracker view previously set for all single edited labels the coordinates of the last edited one

○ Progress for a single uploaded photo did not change, causing the system to block entry into the dataset with the completed Upload



14.04.2022 - 08.07.2022

  • Added:
    • OSAI:
  • Introduction of a tutorial mode for inexperienced users with the option to enable and disable it from the user settings
  • Introduced user configuration (tutorial settings and 'dark mode')
  • Pytorch framework added and classification learning implemented
  • Added Object tracker - i.e. the ability to upload images from a camera connected to the user's computer with the function of tracking the selected object
  • "Predictor" functionality has been added to the Dataset window's detail view. The user can select a previously trained model and then run a prediction on a given dataset.
  • Expansion of the dashboard - section collapsing, division of tables into active and recently modified. Added "Working time" field, with current time in active tasks.
  • Conversion and prediction sections were added to the dashboard view
  • AIDWS view redesign. Added navigation bar along the lines of the one from OSAI with Cameras, Live testing, Results and Documentation tabs.
  • Added a view to add, remove and edit USB and CSI cameras connected to the user's device with full compatibility with AICS
  • Added ability to enable live viewing of USB and CSI cameras connected to the user's device
  • Added creation of "Datasets" based on streaming results to the main portal, from sources such as photos, usb camera, or video files with and without model overlay
  • Expanded the dataset to include a list of supported USB cameras and CSI, as well as default configurations for these cameras
  • Support for the new version of OpenVino in AICS


  • General:


  • Added a new worker to create predictions and send them to the backend
  • A docker image for the manipulator was created. Adaptation of the manipulator to operate inside the system in the same way as other workers (predictor, trainer, converter)
  • Ability to add a custom device within Live Testing, docker installation instructions have been provided for the device to connect to OSAI
  • Updated:

○ In the "Upload dataset" view, when you click "Upload", a window is displayed asking for the action to be taken for images that do not meet the requirement (scaling or skipping)

○ In the "Upload dataset" view, an option to collapse the empty image preview (Preview) has been added

○ Improved error handling from the backend-frontend. Increased the number of messages displayed to the user for specific errors

○ Rebuilt incoming statuses from baggers, increased number of supported stages, included startup and shutdown of docker, increased responsiveness to status changes on the frontend side

○ Supported breaking websocket connection. Recovering the connection refreshes the statuses of active tasks on the dashboard

○ Training results - statistics - have been moved from the database to S3 Storage

○ To close the "Does not exist" message, click it - previously it disappeared after 5 seconds.

○ Unification of the fields of workers - frontend

○ Adding the caption "Preview of the data to upload" in the dataset preload view

○ Adding a status represented by a green circle indicating the correct execution of an action in labeling

○ Adding a spinner in the "Results from subset" section of the training detail view

○ When preloading the dataset on the page (Preview), added a window asking the user if he/she is sure he/she wants to leave the dataset loading process with the caveat that in this case his/her images will not be uploaded to the server

○ Handling prediction results in reports

○ Improved colors on the page to increase clarity and improve the overall user experience in 'dark mode'

○ Percentages on the tile in the model listview first retrieve the percentage value from the database, and then are updated in real time when statuses arrive from the backend

○ Improved responsiveness of confusion matrix table in model detail view

○ Mandating selection of a specific operation in the manipulator view before adding another one

  • Fixed:

○ No validation in the name of the manipulated dataset

○ Renaming the "Send" button to "Start manipulation" in the manipulator dataset view

○ Blocking the addition of empty manipulations

○ Improved percentage results in employee tables on the dashboard

○ Now in the model details view on the charts, the number of epochs starts at 1 instead of 0

○ Fixed errors causing the manipulator to close

○ Centered the caption "no data" when there is no data.

○ Fixed the positioning of the "version" caption containing the current version number

○ Margins between tables on the dashboard

○ Error with floating point number in the learning rate field - training parameterization

○ Improved styling of the "Category Manager" view.

○ Support for invalid category names in model configuration

○ Collapsing the "Back" button when the side menu is collapsed



06.04.2022 - 14.04.2022

  • Added:

○ After adding a manipulation, we inform the user about the queue position in the dashboard view.

○ When a workout is added, we inform the user of the queue position in the dashboard view.

○ When merging categories, a message has been added inside an empty bucket that it should be deleted or categories should be added

○ It is now possible to download mistakenly uploaded files as a csv file

○ Added the ability to view the page using a darker color palette (Dark mode)

○ Progress of downloading models on AIDWS

  • Updated:

○ Unified drawing of inference results on Jetson and VM

○ Upgrade of .NET 5.0 to .NET 6.0 on AIDWS

○ Now the behavior of upload, training and manipulation on dashboard is consistent (you can see the finished task, you can enter the record)

○ In shuffles now we show integer values

○ We receive confirmation that the task has been chained. Otherwise, the task is deleted and the user is notified of the error.

  • Fixed:

○ In rental devices, it no longer shows the status "Creating inference engine" when loading an existing model

○ Now the default accuracy value is correctly shown during automatic training configuration

○ No server cleanup mechanisms when image upload is interrupted

○ Bug when converting old models to Nvidia Maxwell

○ Fixed translation errors on AIDWS

○ Model download cancellation did not interrupt backend uploads

○ Fixed duplicate shuffles

○ Fixed non-expanding shuffles

○ Fixed navigation not working when you want to move from one dataset to another or from one training to another via notification

○ Fixed bug with missing expected Accuracy for automatic configuration



23.02.2022 - 06.04.2022

  • Added:

○ Possibility to rent a virtual machine with support for models from Keras and Darknet frameworks

○ Ability to convert models to ONNX

○ New view for adding models, along with a feature to automatically configure training by indicating only one parameter - the expected accuracy of the model,

○ Consistency in buttons - blue color for main actions, ghost button for side actions

○ Notification of incorrectly uploaded images in dataset after upload

○ Adding icons in Back navigation buttons i.e. consistency - each back button has a back arrow icon that suggests going to the previous window.

○ Editing in the view of adding labs - blocking delete buttons if there are no categories, 'No data' if there are no labs

○ Notifying the user of datasets that do not meet the requirements for detection or classification (category selection view), adding a button to redirect to the dataset view

○ Deletion of multiple images simultaneously in the detail view. Adding confirmation of deletion of photos

○ Deleting a category in the upload view is now with two possible options: Deleting the category itself and keeping the photos belonging to it, or deleting the category along with the assigned photos

○ Spinner when uploading reports

○ Forced password change when accessing AIDWS for the first time

○ Improved stability and performance of database connections

○ Service that removes in the background files inconsistent with the database

○ Adding a new slider to divide dataset into subsets train, test , validation without moving labels + maintained color consistency

○ Selection of datasets and categories before selecting network parameters

○ Category manager allowing to rename all categories and delete unused ones

○ Ability to choose automatic or manual configuration during model creation

○ Type of leased device has been added to the report view. (Important: No backward compatibility. You need to generate a new report to see the device type)

○ A link to the documentation has been added to the main menu

○ Added a warning regarding .zip loading that there may have been problems through link instability

○ Added information not to turn off the card or refresh the page during upload

○ When renting Jetson/VM, added page indicating that machine setup is in progress

○ "Preparing" status added to conversions that are waiting in the queue

  • Updated:

○ Removing the two-step conversion for Jetson Nano. Clicking the conversion "NVIDIA Maxwell" creates the engine. With this solution, the user does not have to wait for the conversion, which took place earlier on the borrowed jetson

○ Preparation of training and manipulation is done on a separate thread, in the background, so the client does not have to wait

○ Detailed view of the model broken into tabs

○ Upgrade from .NET 5.0 to .NET 6.0 including libraries

○ Improved system-wide performance

○ Significant acceleration of photo uploads

○ Implementation of a new library for uploading ZIPs,

○ Changed table pagination to scroll

○ Rebuilt view of adding models and templates, separate configuration components for given frameworks

○ Instead of small windows with delete confirmation, introduced modal in the middle of the screen with delete confirmation - view of datasets, models and templates

○ Accelerated download of datasets in Pascal/Voc and Darknet formats

○ Speeding up the download of recent manipulations in the Dashboard view

  • Fixed:

○ Excessive number of images caused training to break in Tensorflow Keras framework

○ Fixed bugs in processing certain types of images

○ Fixed bug in Darknet parser that caused rejection of valid labels

○ Fixed bugs with locking and unlocking datasets

○ Fixed bugs with uploading ZIPs

○ Fixed release of borrowed devices

○ Remember selected hardware on AIDWS, when returning from a model the application returns to the correct location

○ 'Back to Deployment' button changed to 'Back to Live Testing' button

○ Task counter now also updates during training, etc.

○ Scrolling of notifications and reports now does not jump

○ Improved validation for datasets and model names - use of spaces and removal of spaces from beginning and end

○ Fixed errors when renting devices

○ Loss when entering the model list during training displayed ugly

○ When merging categories when deleting a bucket, it did not rename and interpreted as one bucket

○ Improved the view of buttons on mobile devices - maintaining the width and sticking to the bottom of the screen

○ Warning before refreshing the page or closing it while the upload is in progress - correcting an error resulting from a badly downloaded status

○ Fix bug with lack of progress on the dashboard when uploading a ZIP

○ Locking the save template button when no changes are made when creating models

○ Double dot in statuses from manipulator

○ Bug when quickly adding new labels in labeling view causing misassignment

○ Improved animation in tile view in datasets, models and templates, added animation that also works when moving the mouse outside the tile

○ Cancellation of downloading datasets was not freeing up resources on the backend



08.02.2022 - 22.02.2022

  • Added:

○ Support for annotations in Darknet TXT format without labels.txt file - classes are then given default names, i.e. Class0, Class1, Class2, etc.

○ In the labeling view in the "Details" tab, added image size in KB

○ Full progression when uploading a ZIP file - from the moment the client sends it to the server (the first 20%) to the completion of the task in the background.

○ Automatic repair of erroneous labels on the client side, before displaying them to the user

○ Automatic repair of labels in Darknet TXT format on the server side

○ Progress when preuploading photos on the client side, along with a counter of processed files

○ After uploading images, we have information not only about errors, but also warnings about annotations, for example, when any was automatically fixed or was completely incorrect and was rejected

○ An algorithm for automatically matching annotations to photos based on directory and file name matches, and then the file names themselves. Annotations in Pascal VOC format have higher priority than those in Darknet TXT format

○ In preupload mode, processed photos are annotated with classification categories

○ Error notification when ZIP upload fails, which comes immediately - no need to wait for timeout

○ Added model history in the model details view

○ Tasks in the top bar have a counter added

○ Portal version information in the lower left corner

○ Added training results view in model details

○ Bell animation when notification comes in

○ Added pagination to the table in Training Details for Keras and Darknet models

○ Added check if model name is not occupied during stage 1 retrace

○ Added restriction (browser hint) for files uploaded to AIDWS. Currently supported formats are .jpg,.jpeg,.mp4,.png,.bmp

  • Updated:

○ The gradient on the tile in the manipulator has a different shade

○ Modified view of tab headers in dataset details and uploads

○ Modified view of images in upload mode - they now have the same style as when viewing the uploaded dataset

○ Deletion of inactive tasks now starts once every half hour and notes whether workouts and conversions are queued - these are deleted after 24h

○ Significantly improved file processing performance and linearity of the process - it no longer slows down as the number of files increases

○ Improved performance in uploading thumbnails in photo preupload mode

○ Improved efficiency in server-side annotation validation

○ Unified processing of photos and annotations in direct upload mode and in ZIP file mode

○ After selecting a different number of photos on a page, there is an automatic recalculation of which page we should go to in order to still see (at least partially) the same photos

○ Increased interactivity in the retrain wizard - added hints in category merging informing what needs to be corrected

○ List of models on AIDWS as tiles

○ Video player on AIDWS scales better

○ Improved transition between pages on AIDWS

○ Additional validation of fields in Keras and Darknet trainers

○ Additional validation of fields in DatasetManipulator

○ Significant security improvements in terms of validation of transmitted requests between frontend and backend in all endpoints

○ Percentages in confusion matrix instead of number of images

  • Fixed:

○ Failure to check if the model exists when trying to download it

Infinite loading icon error during preupload when there was a ZIP in the uploaded folder

○ Error during ZIP upload - until the file was uploaded to the backend, the dataset was not locked and the corresponding task did not appear on the portal

○ Errors when dragging files in upload mode

○ Errors when processing annotations on the client side

○ Errors when deleting and editing categories in file upload mode - no more disappearing and unexpectedly returning files, and category names are correctly uploaded to the server

○ Bug allowing to start uploading when some files were left unprocessed and resulting in images without annotations

○ Bug allowing to download a dataset that has no photos

○ Bug allowing to manually enter upload mode despite the fact that the dataset was locked

○ Many minor bugs when uploading images and ZIP files

○ Change of the number of images on the last page in dataset details

○ Number of images after upload refreshes automatically

○ Unable to start object detection with one category

○ Notifications sort after time

○ Delete button locks up when clicked

○ Fixed conversion of models to OpenVino

○ Locked Retrain button when pressed so that it cannot be pressed twice

○ Fixed a bug that allowed the upload button to be pressed when a file was already uploaded - AIDWS