Authors since 5.0.0: Jordi Torrents
Improved branding design with a new logo and icon.
Fix critical bug that was making knowledge transfer crash when tracking videos with different identification image sizes.
Improved Validator stability.
Reallocation of source code files.
Catch crash in MacOS when
BlockingIOErrorraises at opening H5FD files in mode “r+”.
Added resilience against corrupted videos avoiding wrongly decoded frames.
Added advanced hyper-parameters
MAXIMAL_IMAGES_PER_ANIMALin docs and in terminal argument parser.
More intelligent automatic zoom in Validator based on animal’s body length.
Cleaner logging by not printing the level name if it is
Allow float and infinite values in blob’s area and intensity thresholds.
Simplified area widget in Segmentation App.
Define default session’s name by double clicking in the session’s name widget in the Segmentation App. The App will save .toml files without names if the widget is empty.
THRESHOLD_EARLY_STOP_ACCUMULATIONchanged from 99.95% to 99.9%.
idtrackeraiterminal command. Instead, multiple parameters files can be loaded with the
--loadargument in increasing order of priority.
More compact representation of
Simplified intensity thresholds in Segmentation App.
Fixed identity removal in Validator.
list_of_blobs_validated.pickle. Validated blobs are saved in the same original file
Added “Autoselect error” option in the Validator.
Fix crucial error of not properly shuffling images before splitting into train/validation datasets.
Improved memory efficiency.
Automatic session names made from video paths if no session name is provided.
Do not override sessions if their name has been automatically set. Create consecutive names like session_X_1.
Add a test step after each training.
--sizeparameter for individual Video generators.
The default value for
number_of_parallel_workersis limited to 8 maximum.
Do not save list_of_blobs_no_gaps.pickle.
Merge network architectures for crossings and identification.
Simplify video JSON file.
Add boolean flag
identity_transfer_succededto video JSON file.
accumulation_statisticsfrom video JSON file.
Faster training with efficient image normalization.
Faster coexisting Fragment connection.
Faster “First individual/crossing assignment”.
Do not fix identity of small Fragments.
More memory efficient network predictions.
Fix crash in the video generator when using
Add a button to remove the selected centroid when double clicking (Validator).
Allow loading a TOML parameters file with the Open button in the Segmentation app.
More informative logs
Fix invalid model predictions when using Metal backend in MacOS machines polavieja_lab/idtrackerai#82.
Fix Overflow crash when validation loss is exactly 0.
Fix crash of Protocol 3 in Windows because of the default integer type in Numpy.
LEARNING_PERCENTAGE_DIFFERENCE_*hyperparameter renamed to
Allow idtrackerai to keep working even if OpenCV fails reading some video frames.
Limit framerate option in GUI enabled by default.
number_of_parallel_workers=1disables Python’s Multiprocessing.
Fix video generator when dealing with error frames.
background_subtraction_statto Segmentation App.
More informative logs, specially in the accumulation results.
Catch exception when it fails to read the number of frames of a video.
Fix NumPy integer types.
Add defaults values in
Improve errors and tracebacks in log.
Works in Python 3.10 and 3.11.
Improve error messages.
New option to add a time column (in seconds) in the csv trajectory files. Parameter
CONVERT_TRAJECTORIES_TO_CSV_AND_JSONdefault changed to
Reorganize trajectories output folder.
Change video extension limitation for everything OpenCV can read.
output_direrror when it is stated in toml file.
Fix Protocol 3 with knowledge transfer
Merge and simplify
More stable Validator with unfinished sessions.
Clearer code in entry point functions and parameter management.
Fix Validator bugs
Zoom un duplicates when clicking this error in Validator.
Setup points as integers
Fix input parameters effect on segmentation GUI
Remove deprecated image blurring parameter
Blob.is_an_individualsetting when crossing detection training fails.
Fix distorted image visualization in some video formats.
Fix crash due to the missing
Fragment.P1_vectorattribute while validating.
Abstract Qt dependencies with
Default Qt package downgraded from PyQt6 to PyQt5 because its compatibility issues.
Allow running non-GUI idtrackerai parts without any Qt installation.
Add identity finder in Validator with Ctrl+F
More versatile command line options for
(testing) Allowing running idtrackerai in CPU only mode, AMD GPUs and MacOS with or without MPS acceleration.
ListOfGlobalFragmentsare saved in .json format.
ListOfFragmentsare saved in .json format.
Load the penultimate accumulation step if the last one broke.
Focus on error when selecting an “No id” error in Validator.
Reenable blobs’ contours approximation using less points (lighter blobs objects in RAM and disk).
Speed up trining by enabling
Improve Validator responsiveness when loading a large session.
Remove scikit-learn dependency.
Fix GUI initialization error in Fedora
Fix background view in Segmentation App.
Improve logging information.
Fix final compression bug on Windows.
idtrackerai_videoincompatibility when tracking without identities.
Fix GUI theme change malfunctioning.
Disables the first two changes of changelog 5.1.1 (the identification images construction method and the blob’s contour approximation). These will be restored after some more testing.
Maximize the number of relevant pixels inside identification images (faster identification).
Approximate blobs’ contours using less points (lighter blobs objects in RAM and disk)
Allow a variable number of animals (by setting n_animals=0) when tracking without identities.
Removed anti-flickering filter. It improves intensity threshold’s sensitivity and segmentation speed.
Using gzip compression on identification images files when finishing a successful session. Optimizing loading times.
Added playback speed in a action in the video player menu.
Python’s datetime usage in Video timers
Free Enter/Return keys from GUI shortcuts.
Fix log file issue when more than one session is running.
Optimize rescaling when resolution reduction and adapt ROI to scale.
Fix bug when validating single animal trackings.
Fix “change font size” bug in GUIs.
Ctrl+L to toggle playback framerate limitation (GUI).
Fix cv2 BRG/RGB color confusion.
Fix cv2 error in segmentation app when removing ROI while using background subtraction.
ROIs can be reordered in segmentation app by drag and drop
ROIs in segmentation app are ordered from bottom to top
Select ROI by clicking inside the polygon in the video player (segmentation app)
Simplify idtrackerai/network file structure and imports
Improve v4 compatibility reading video.json/npy
Merged crossings/identification NetworkParams as a single dataclass
Simplified GetPredictionIdentification converting it into a function
Fix gray individual video generation
Full code revision promoting Python built-in libraries, argument type hints and multiples optimizations in terms of code simplicity and structure, RAM usage, lighter output generated data and faster execution.
Unify all tools related to idtracker.ai in the same repository/package
Works with Python 3.10
New graphical apps buildings directly with PyQt6
Remove dependency with
Using the last versions of every remaining used dependency
Completely new segmentation app, fully responsive, intuitive and faster
Completely new validation app to view the session results, navigate through the possible tracking errors, fix them and manipulate the session using other extra tools.
idtrackerai_csvtool to convert trajectories after the tracking process finished.
Removed local_settings input method.
New input methods (more direct and simple) (
--settingsand terminal declarations).
The tool “setup points” moved to the validator.
NUMBER_OF_JOBS_FOR_BACKGROUND_SUBTRACTION, background is computed sequentially.
NUMBER_OF_JOBS_FOR_SETTING_ID_IMAGESin the same parameter
Easier background subtraction implementation, with “median” option. It is more robust against difficult tracking intervals/episodes/number of frames.
Better and easier parallel episode definitions with optimized parallel distribution (specially with multiple files).
Simplified attributes in all idtracker.ai objects.
ListOfBlobs reconnects in almost no time after loading from saved .pickle file.
Flexibility selecting the number of videos to track.
Remove Blob.pixels attribute. Much faster and lighter blob manipulations.
Stretch Blob.bounding_box. Much lighter segmentation images.
Optimized 80% of the computational time of _process_frame() by properly removing the function binary_fill_holes().
Logs more readable, with more useful information and progress bars (using Rich).
Faster h5py writing/reading implementation (by not opening and closing the h5py file for every single image, we keep them opened).
Python objects are saved as pickle objects and json files when possible (lighter and more standard than .npy files).
Removed option save_areas. Now, the statistics of the areas are always printed in the trajectory files.
Parallel processing using built-in Multiprocessing, not Joblib.
Reorganize internal modules promoting decoupling (fragmentation, tracking and postprocessing modules).
Easy video generation with
Package is defined using a pyproject.toml file.
No git sub-modules used.
Faster blob overlapping method (convexHull and point inside contour methods).
Works with Python 3.7.
Remove Kivy submodules and stop support for old Kivy GUI.
Neural network training is done with Pytorch 1.10.0.
Identification images are saved as uint 8.
Crossing detector images are the same as the identification images. This saves computing time and makes the process of generating the images faster.
Improve data pipeline for the crossing detector.
Parallel saving and loading of identification images (only for Linux)
Simplify code for connecting blobs from frame to frame.
Remove unnecessary execution of the blobs connection algorithm.
Background subtraction considers the ROI
Allows to save trajectories as csv with the advanced parameter CONVERT_TRAJECTORIES_DICT_TO_CSV_AND_JSON (using the local_settings.py file).
Allows to change the output width (and height) of the individual-centered videos with the advanced parameter INDIVIDUAL_VIDEO_WIDTH_HEIGHT (using the local_settings.py file).
Horizontal layout for graphical user interface (GUI). This layout can be deactivated using the local_settings.py setting NEW_GUI_LAYOUT=False.
Width and height of GUI can be changed using the local_settings.py using the GUI_MINIMUM_HEIGHT and GUI_MINIMUM_WIDTH variables.
Add ground truth button to validation GUI.
Added “Add setup points” featrue to store landmark points in the video frame that will be stored in the trajectories.npy and trajectories_wo_gaps.npy in the key setup_poitns. Users can use this points to perform behavioural analysis that requires landmarks of the experimental setup.
Improved code formatting using the black formatter.
Better factorization of the TrackerApi.
Some bugs fixed.
Better documentation of main idtracker.ai objects (video, blob, list_of_blobs, fragment, list_of_fragments, global_fragment and list_of_global_fragments).
Dropped support for MacOS