Contributor

Implements the creation of ReID transformer dataset creation for PyTorch pose estimation models.

• implements deeplabcut.create_tracking_dataset for PyTorch models
  • Closes np.nanmean bug in creating triplets for transformer re-ID: query_feature_by_coord_in_img_space() function #2794
  • Closes Unsupervised ID tracking - This function is not implemented for Engine.PYTORCH #2809
  • Creates a FeatureShelfWriter to save ReID features

Other changes

• removes the dynamic parameter from deeplabcut.create_tracking_dataset as it cannot be used there
• adds a docstring for deeplabcut.create_tracking_dataset
• correctly pulls the pcutoff from the inference config when creating assemblies in convert_detections2tracklets
• adds an image_size key to the context when running inference, and a top_down_crop_size for top-down models

base code for reID tracker

ea87ce4

added postprocessor tests

0e51357

apply fix from #2794

40b3e97

improved docs

aa6e8ae

fix preprocessor creation for top-down models

7e217b0

Merge branch 'pytorch_dlc' into niels/reid

5a6fed4

use pcutoff from inference_cfg when creating tracklets

a963df6

fix dealing with NaNs

21b0310

update userguide

a31e9d8

add missing docstring

6ac9468

add missing docstring

86da41c

maximpavliv approved these changes Feb 7, 2025

View reviewed changes

Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, thanks @n-poulsen !

np.nanmean bug in creating triplets for transformer re-ID: query_feature_by_coord_in_img_space() function #2794

Unsupervised ID tracking - This function is not implemented for Engine.PYTORCH #2809

Fix: properly use unique_bodyparts and identity_scores arguments in ShelfWriter.add_prediction #3112