Skip to content

changelog.rst

Latest commit

 

History

History
608 lines (514 loc) · 32.5 KB

changelog.rst

File metadata and controls

608 lines (514 loc) · 32.5 KB
Raw

Changelog

Read about the new identification algorithm in :ref:`6.0.0` in our latest publication

idtracker.ai publication in eLife (2026) idtracker.ai publication in eLife (2026) 
.. centered:: `New idtracker.ai: rethinking multi-animal tracking as a representation learning problem to increase accuracy and reduce tracking times <https://doi.org/10.7554/eLife.107602>`_

Authors since :ref:`5.0.0`: Jordi Torrents (jordi.torrents@research.fchampalimaud.org | jordi.torrentsm@gmail.com)

idtrackerai-detectron2 (this fork)

Changes made in this fork and not present in upstream idtracker.ai. They are described at length in VALIDATOR.md and README.md in the repository root, and in docs/detectron2-pipeline.md and docs/sam3-pipeline.md.

Validator changes

  • New Swap with..., Extend... and Merge with... operations in the double-click dialog, backed by new Blob.swap_identity(), Blob.propagate_swap_identity() and ListOfBlobs.merge_blobs().
  • Undo the last twenty edits with Ctrl+Z.
  • A failed save now writes the session to ~/idtrackerai_backups and reports it, rather than losing the work silently on a worker thread.
  • Autosave every five minutes, without progress dialogs.
  • New presence_intervals session parameter: frames in which an identity is not expected to exist stop being reported as missing-identity errors.
  • Minimum-duration and identity-pair filters on the list of errors, and a next-error button bound to N.
  • New PCHIP interpolation mode, which cannot overshoot between fixed points. The default interpolation order is now linear. Enter applies an interpolation, I starts one, and clicking a fixed point jumps to its frame.
  • The length calibration line now follows the cursor. Fixed deleting the wrong calibration when two of them read the same, and fixed a calibration restored from disk raising AttributeError.
  • Holding a frame-step key scrubs through the video instead of stepping once.
  • New Video Info tab.

Segmentation changes

  • New external_contours session parameter: animal outlines are read from a sidecar file produced by an external instance-segmentation model instead of being found by thresholding. It takes one file per video path, so a recording saved in several clips is tracked as one session with identities carried across the joins. The Segmentation App pairs files to clips by name, which makes the order right by construction, and refuses rather than tracking part of a session when a clip has no file.
  • The Segmentation App gains a segmentation-source selector and a guided panel for sampling frames, annotating them and building a Detectron2 dataset.
  • SAM 3 can produce the contours instead of a trained Mask R-CNN, prompted with a word for the animal rather than trained on annotated frames. It writes the same sidecar files, so nothing downstream of segmentation distinguishes the two, and the contour exporter now dispatches on --backend. SAM 3 runs through Meta's sam3 package, which like detectron2 is optional and not declared as a dependency.
  • New idtrackerai_sam3_prelabel: SAM 3 drafts LabelMe annotations over the sampled frames so preparing a Detectron2 dataset becomes correcting outlines rather than drawing them. It never overwrites an existing annotation, writes no file for a frame it found nothing in, and simplifies outlines so they can be edited by hand.
  • The Segmentation App gains a fourth segmentation source, SAM 3, offering both of those. Exporting a whole video is offered only where there is a CUDA GPU, since on a CPU it is not slow but impractical; the Colab notebook covers that case, in a new section 8.
  • export_video() takes progress and abort callbacks. An aborted export writes nothing, because a partial sidecar would be counted as a finished one by the resume logic.
  • The preparation panel keeps its own list of videos to draw annotation frames from, separate from the videos being tracked, so one model can be trained across every recording from a setup while each recording is tracked on its own. It shows the per-video allocation before sampling and warns when the frame count is too low for every video to be represented.
  • The train/validation split now groups clips by recording rather than by file, folding pieces such as trial_segment_1/2/3 together. The inferred grouping is shown before the dataset is built and can be overridden, with --group-by on the command line.
  • Fixed sampled frames recording their source video inconsistently, which let frames from one clip be counted as two sources and placed on both sides of the train/validation split. Re-sampling into a folder now adds to its manifest instead of replacing it.
  • Fixed the frame budget being concentrated on one clip rather than spread across them, and a single unreadable file aborting a whole sampling run. Clips whose file names would collide are now refused before anything is written.
  • Training and contour export are steps in the Segmentation App, not only instructions to go elsewhere. The panel checks whether this machine has a CUDA GPU and Detectron2; when it does, training runs here as a managed subprocess with its output shown and a stop button, without importing Detectron2 into the GUI process. The export is handed over as a command, because roughly an hour per 30 000-frame clip means days over a collection. Colab is now the fallback for machines without a GPU rather than the route.
  • The Colab notebook carries a project from the annotated frames to the end: it builds the COCO dataset, trains, exports one contour file per clip, and writes one idtracker.ai parameter file per recording. Those files inherit the configuration saved by the Segmentation App, so the animal count, area thresholds, region of interest and tracking interval carry over rather than being retyped. Its sections are numbered 4 to 12, continuing from the app's steps 1 to 3 -- enhancement, sampling and annotating, which stay there because each needs eyes on the footage, and because LabelMe cannot run on a runtime with no display. A section that tracks on the runtime is included and marked as untested there.
  • New Get Colab bundle... button on the annotate step, which writes the bundle and lists what else to put on Drive.
  • New idtrackerai_d2_install_gpu command, and a button in step 5, that installs PyTorch and Detectron2 for the machine they run on: the GPU and its CUDA version are detected, the matching PyTorch index is chosen and checked to exist, and the commands are shown before anything runs. Neither package can be an ordinary dependency -- PyPI's PyTorch wheel for Windows is CPU-only, so declaring it would silently disable the GPU for idtracker.ai's own identity tracking, and Detectron2 is not on PyPI at all and builds from source, so declaring it would make installation fail on any machine without a compiler.
  • The Segmentation App's left column scrolls. It asks for more height than a laptop screen has, and the layout answered that by shrinking every widget below its natural size, which left a preparation step too short to show its own button.
  • Fixed the Colab notebook aborting on every freshly built bundle: its integrity check looked for tools/frame_preprocessing.py, a name the bundle has never written. Repaired its markdown cells, whose stored lines had lost their line endings so each cell rendered as one run-on paragraph, and updated the script names the fork has since replaced with installed idtrackerai_d2_* commands.
  • New enhancement session parameter. Frame enhancement (CLAHE and correction for uneven lighting) is chosen against the footage in the Segmentation App, with a live preview and named presets, and applies in every segmentation mode: to the frames, to the background model and to the identification images alike. It is off by default. Turning it on shifts the brightness range, so intensity thresholds chosen against raw frames need re-tuning; the app says so next to the thresholds.
  • The pipeline that produces the contour files ships inside the package, as idtrackerai.extra_tools.detectron2_pipeline, so the GUI and the command line run the same code. Every stage is also an installed idtrackerai_d2_* command.

Documentation changes

  • New USER_GUIDE.md: an illustrated guide for someone who has not tracked an animal before, covering the vocabulary, installation, a first run by thresholding, what its failure looks like on real footage, the Detectron2 route both locally and in Colab, tracking from external contours, and the Validator.
  • New docs/make_screenshots.py, which drives the Segmentation App offscreen to regenerate every screenshot in that guide, so the images can be refreshed when the interface changes.

Packaging changes

  • Distributed as idtrackerai-detectron2. The import name is unchanged, so it cannot be installed alongside upstream idtracker.ai.
  • Fixed the version lookup failing under the fork's distribution name, which made the package raise at import time once installed, and made every tracking run raise before it started.

6.0.15 (unreleased)

  • Enforce utf-8 in all text I/O operations.

6.0.14

  • Fix non working App icon in Linux taskbar (Ubuntu, Gnome).
  • Fix bug when setting custom background in Segmentation App.
  • Allow disabling impossible jumps detection in Validator.
  • Handle PermissionError when tracking from a directory with no write permissions.
  • Add CSV tidy output format (https://gitlab.com/polavieja_lab/idtrackerai/-/work_items/108)

6.0.13

  • Add the entrypoint idtrackerai_background to compute the background of a video or set of videos from terminal, see :ref:`background subtraction`.
  • Remove parameter IDENTITY_TRANSFER setting it to True if a knowledge transfer folder is provided.
  • Fix bug in the :ref:`validator_reference` when saving session from the pop up dialog at closing the GUI.
  • Fix bug in the :ref:`validator_reference` when opening a session from the pop up dialog at starting the GUI.
  • Add identities to fragments when tracking single animal or single fragment videos. This allows to use idmatcher.ai afterwards.
  • Handle sessions without identification models and fragments without identities (when n_animals=1) in :ref:`idmatcher.ai`.
  • Add a "save" button in the Segmentation App to store the computed background as an image file.
  • Do not set the automatic resolution reduction if the recommended value is too close to 1 (bigger than 0.91).
  • Add Parquet format to the possible trajectory output formats, see :ref:`formats`.

6.0.12

6.0.11

6.0.10

  • Add the possibility to upload a custom background in the Segmentation App and by using the parameter BACKGROUND_SUBTRACTION_STAT, see :ref:`Background subtraction`.
  • Ask for a session to open if no session directory is provided when opening the :ref:`validator_reference`.

6.0.8

  • Add color settings in the :ref:`validator_reference` and :ref:`video generator`.
  • Fix incompatibility with older versions of OpenCV.
  • Automatic pop ups in GUIs only when there is a major upgrade available (not in patch updates).
  • Remove bias parameter from the last fully connected layer in ResNet models since only relative distances in the representation space matter.
  • Redefined fragment_connectivity with a factor x2 and added its value to the :ref:`trajectory files`.
  • Removed estimated_accuracy_identified and estimated_accuracy_after_interpolation from the :ref:`trajectory files`. The value of estimated_accuracy should be used instead to estimate the accuracy of the tracking.

6.0.7

  • Fix incompatibility with PyQt5.

6.0.6

  • Implement tracking without identities together with exclusive ROIs. This allows to track videos without keeping track of the identities but still respecting the exclusive ROIs defined in the segmentation app.
  • Fix a bug in the :ref:`validator_reference` with numpy.bool getting into PyQt system.
  • Improve main Argument Parser descriptions.
  • Fix incompatibility with non-ASCII (chinese) characters in paths.
  • Fix frozen Validator GUI when opening and saving sessions due to bad parallel processing management.
  • Add new parameter CONTRASTIVE_MIN_ACCUMULATION to set the minimum fraction of images that need to be accumulated for taking the contrastive step as sufficient. If the fraction of accumulated images is lower than this value, the accumulation protocol will be run. This parameter is set to 0.5 by default.
  • Optimized object rendering in the :ref:`validator_reference` preview by checking if the object is visible in the current frame before rendering it.
  • Fix video player not respecting original video framerate.

6.0.5

6.0.3

  • Fix a critical bug in the Validator that was causing the GUI to crash when exploring the last frames of the video.
  • Fix documentation issues.

6.0.2

  • Hotfix for a critical AttributeError at compute_identity_probabilities.

6.0.1

  • Hotfix for non responding GUI issues on MacOS.

6.0.0

  • The Protocol 3 (used as a fallback of Protocol 2) has been removed and Contrastive Protocol has been introduced as the new main identification protocol leaving Accumulation Protocol (Protocol 2) as a fallback to run if Contrastive fails (publication in progress).
  • Trajectories can be saved in a variety of formats following the new parameter TRAJECTORIES_FORMATS (:ref:`check the docs <output>`). As a consequence, the parameter CONVERT_TRAJECTORIES_TO_CSV_AND_JSON has been removed.
  • Resolution reduction has been removed from the Segmentation App and it has now a default automatic value, see :ref:`knowledge transfer`.
  • The different trajectories files have been merged into one single file. There are no longer trajectories_with_gaps, trajectories_wo_gaps nor trajectories_validated, only trajectories. This single file acts as trajectories_wo_gaps and it contains all modifications from the :ref:`validator_reference`.
  • Since Protocol 3 does not exist anymore, the following parameters have been removed:
    • PROTOCOL3_ACTION
    • THRESHOLD_ACCEPTABLE_ACCUMULATION
    • MAXIMUM_NUMBER_OF_PARACHUTE_ACCUMULATIONS
    • MAX_RATIO_OF_PRETRAINED_IMAGES
  • Remove the parameter ADD_TIME_COLUMN_TO_CSV setting it to always True.
  • Added property :attr:`~.Blob.identity_certainty` to :ref:`Blob` and used it to populate id_probabilities in :ref:`output structure`.
  • Added automatic :ref:`Code reference` of main classes in documentation.
  • Refactored and optimized :ref:`idmatcher.ai`.
  • Added the new parameter TORCH_COMPILE to enable model compilation with torch.compile, see :ref:`Advanced parameters`.
  • Added the option to respect or ignore the tracking intervals while looking for errors in the :ref:`validator_reference`.
  • Fixed GUI compatibility issues with PySide6.
  • Works with PyQt6
  • Allow files drag and drop into the GUIs to open videos, parameters files or sessions.
  • Use gzip compression in trajectory HDF5 files.
  • Rename time column in CSV trajectory file from seconds to time.
  • Rename MINIMUM_NUMBER_OF_FRAMES_TO_BE_A_CANDIDATE_FOR_ACCUMULATION to MIN_N_FRAMES_TO_BE_A_CANDIDATE_FOR_ACCUMULATION.
  • Added height, width and silhouette_score to :ref:`trajectory files`.
  • Fix Validator bug when acting on a centroid that has a duplicated identity.
  • Fix and speed up blob-to-blob overlapping check.
  • The default value for number_of_parallel_workers is limited to maximum 4 since the parallelized tasks involve disk read and write operations that are not CPU bound.
  • Added automatic usage analytics, see :ref:`Usage Analytics`.
  • Hardcoded LEARNING_RATE_IDCNN_ACCUMULATION to 0.005.
  • Default Multiprocessing start method to forkserver when possible.
  • Added a new frame preloader in the background of the Validator to improve video playback speed when browsing the list of errors.
  • Improved GUI's Video players responsiveness, specially when dealing with heavy-loading video files.

The contrastive algorithm design for this version was developed by Jordi Torrents, Tiago Costa and Gonzalo G. de Polavieja and published in eLife (2026).

5.2.12

  • Works in Python 3.12, 3.11, and 3.10. And NumPy 2.0.
  • Increase and decrease GUI font size with Ctrl++ and Ctrl+-.
  • Fix https://gitlab.com/polavieja_lab/idtrackerai/-/issues/90
  • Optimized "Connecting coexisting fragments".
  • The log file copy in session folder contains error tracebacks.
  • Add DEVICE as an optional input parameter.
  • Fix bug occurring when session folder contains non-ASCII characters.
  • Add --no-labels parameter to :ref:`Video generator` (https://gitlab.com/polavieja_lab/idtrackerai/-/merge_requests/73 by https://gitlab.com/ssfrz)
  • Another fix to "Too many open files" error by disabling pin_memory in Protocol 3 pre-training.
  • Fix the automatic gap detector in Validator's interpolator when starting the interpolation by double clicking on a centroid.
  • Fix individual videos cropping to keep the centroid of the animal in the center of the video when some padding has to be added.

5.2.11

  • Fix critical bug making GUI_tools/icon.svg not be included in package build.
  • Set pin_memory=False to non-training DataLoader to avoid "Too many open files" error, pytorch/pytorch#91252.
  • Limit the number of images per animal to use in :ref:`idmatcher.ai` to 10k.

5.2.10

  • Added a :ref:`length calibration` tool in the :ref:`validator_reference` and its value length_unit in the trajectory files.
  • Added the parameter bounding_box_images_in_ram to avoid saving bounding box images on disk.
  • Added https://gitlab.com/polavieja_lab/midline to :ref:`data analysis`.
  • Refactored tracking agent code and merged Protocol 1 into Protocol 2 (no effect on the algorithm).
  • Cleaned Session.accumulation_folder attributes.
  • Add color shuffling action in Validator.
  • Validator interpolates now with splines using scipy.interpolate.make_interp_spline. Previous backend scipy.interpolate.interp1d is now considered legacy by SciPy.

5.2.7

  • 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.
  • Change default DATA_POLICY from "all" to "idmatcher.ai".
  • Improved :ref:`validator_reference` stability.
  • Reallocation of source code files.
  • Catch crash in MacOS when BlockingIOError raises at opening H5FD files in mode "r+".

5.2.6

  • Added resilience against corrupted videos avoiding wrongly decoded frames.
  • Added advanced hyper-parameters THRESHOLD_EARLY_STOP_ACCUMULATION, THRESHOLD_ACCEPTABLE_ACCUMULATION and MAXIMAL_IMAGES_PER_ANIMAL in 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 DEBUG or INFO.
  • 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_ACCUMULATION changed from 99.95% to 99.9%.
  • Removed --settings argument from idtrackerai terminal command. Instead, multiple parameters files can be loaded with the --load argument in increasing order of priority.
  • More compact representation of list_of_fragments.json.
  • Merged hyperparameters BATCH_SIZE_PREDICTIONS_IDCNN and BATCH_SIZE_PREDICTIONS_DCD into BATCH_SIZE_PREDICTIONS
  • Hardcoded LEARNING_RATE_IDCNN_ACCUMULATION to 0.005.

5.2.5

  • Simplified intensity thresholds in Segmentation App.
  • Fixed identity removal in Validator.
  • Deprecate list_of_blobs_validated.pickle. Validated blobs are saved in the same original file list_of_blobs.pickle
  • Added "Autoselect error" option in the Validator.
  • Fix crucial error of not properly shuffling images before splitting into train/validation datasets.

5.2.4

  • Improved memory efficiency.
  • Removed IDCNN_NETWORK_NAME hyperparameter.
  • Rename class Video to Session and video_object.json to session.json.
  • Rename parameter session to name.
  • 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.

5.2.2

  • Add --size parameter for individual :ref:`video generator`.
  • The default value for number_of_parallel_workers is 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_succeded to video JSON file.
  • Remove accumulation_statistics from 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.

5.2.1

  • Fix crash in the video generator when using --gray.

5.2.0

  • Add :ref:`exclusive regions of interest` new feature (https://gitlab.com/polavieja_lab/idtrackerai/-/merge_requests/58).
  • Add a button to remove the selected centroid when double clicking (:ref:`validator_reference`).
  • Allow loading a TOML parameters file with the Open button in the :ref:`segmentation app`.
  • More informative logs
  • Fix invalid model predictions when using Metal backend in MacOS machines https://gitlab.com/polavieja_lab/idtrackerai/-/issues/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.
  • Rename CustomError for IdtrackeraiError.
  • LEARNING_PERCENTAGE_DIFFERENCE_* hyperparameter renamed to LEARNING_RATIO_DIFFERENCE_*

5.1.9

  • 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=1 disables Python's Multiprocessing.
  • Fix video generator when dealing with error frames.
  • Add background_subtraction_stat to Segmentation App.
  • More informative logs, specially in the accumulation results.
  • Catch exception when it fails to read the number of frames of a video.
  • Lighter ListOfBlobs and ListOfFragments files, cleaning cached_property before saving.

5.1.8

  • Fix NumPy integer types.
  • Add defaults values in video_object.json.
  • Improve errors and tracebacks in log.

5.1.7

  • 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 ADD_TIME_COLUMN_TO_CSV (False by default).
  • CONVERT_TRAJECTORIES_TO_CSV_AND_JSON default changed to True.
  • Reorganize trajectories output folder.
  • Change video extension limitation for everything OpenCV can read.
  • Fix output_dir error when it is stated in toml file.
  • Fix Protocol 3 with knowledge transfer
  • Merge and simplify learning_percentage_difference hyper-parameter
  • More stable Validator with unfinished sessions.
  • Clearer code in entry point functions and parameter management.

5.1.6

  • 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

5.1.5

  • Fix Blob.is_an_individual setting when crossing detection training fails.
  • Fix distorted image visualization in some video formats.
  • Fix crash due to the missing Fragment.P1_vector attribute while validating.
  • Abstract Qt dependencies with qtpy package.
  • 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 idtrackerai_csv
  • (testing) Allowing running idtrackerai in CPU only mode, AMD GPUs and MacOS with or without MPS acceleration.

5.1.4

  • ListOfGlobalFragments are saved in .json format.
  • ListOfFragments are 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 persistent_workers=True in torch.utils.data.DataLoader.
  • 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.

5.1.3

  • Fix final compression bug on Windows.
  • Fix idtrackerai_video incompatibility when tracking without identities.
  • Fix GUI theme change malfunctioning.

5.1.2

  • Disables the first two changes of changelog :ref:`5.1.1` (the identification images construction method and the blob's contour approximation). These will be restored after some more testing.

5.1.1

  • 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.

5.1.0

  • Implement idmatcher.ai
  • 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)
  • Fix typos
  • 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

5.0.0

  • 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
    • Pyforms
    • Python-video-annotator
    • Matplotlib
    • Joblib
    • Natsort
    • Tqdm
    • Pandas
    • Gdown
  • 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_csv tool to convert trajectories after the tracking process finished.
  • Removed local_settings input method.
  • New input methods (more direct and simple) (--load, --settings and terminal declarations).
  • The tool "setup points" moved to the validator.
  • No NUMBER_OF_JOBS_FOR_BACKGROUND_SUBTRACTION, background is computed sequentially.
  • Merged NUMBER_OF_JOBS_FOR_SEGMENTATION and NUMBER_OF_JOBS_FOR_SETTING_ID_IMAGES in the same parameter NUMBER_OF_PARALLEL_WORKERS.
  • 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 idtrackerai_video.
  • Package is defined using a pyproject.toml file.
  • No git sub-modules used.
  • Faster blob overlapping method (convexHull and point inside contour methods).

4.0.0

  • 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