Release v10.0.33

• Detect number of classes for clip models automatically based on weights name:

  • 'market1501': 751
  • 'duke': 702
  • 'veri': 576
  • 'vehicleid': 576
  • Previously the num classes was fixed to 751 (market1501) making all the other model loadings crash

• Custom yolov8 models can now be loaded

• Added imgsz to yolo arguments

Release v10.0.32

Yanked. No updates

Release v10.0.31

• ReID crops processing update: from Torchvision transforms to cv2 + numpy operations for full controll
• Feature extraction for crops moved from trackers into ReIDMultiBackend (https://github.com/mikel-brostrom/yolo_tracking/blob/0028c066d700f5e60d456c16a68860f58e3a8838/boxmot/appearance/reid_multibackend.py#L297C12-L304)
• Optimize detections feature extraction
• Adapt tracking options with appearance to all these changes
• Added missing __init__.py to lmbn folder

Release v10.0.30

• Move save path from EXAMPLES (yolo_tracking/examples/runs) to ROOT (yolo_tracking/runs)
• Videos / frames and .txt results are now saved to the same folder
• Classes passed from val to track
• Fixed file not found in colab due to modified save_dir
• Fixed colab after all these modifications

Release v10.0.29

• Added warmup for ReID models
• StrongSORT optimization (minimized bbox conversions) & cleanup

Release v10.0.28

• Added LICENSE banner to all .py files
• Deleted unused backbones

Release v10.0.27

• Fix dynamic input TFLite export: PINTO0309/onnx2tf#441
• Test reid exports: torchscript, openvino, onnx and tflite
• Set default values for bytetrack

Release v10.0.26

Fix warnings that arised from deprecated ultralytics imports

Release v10.0.25

• Modify detectors in order to make them comply with ultralytics' models
• Clean up track.py by letting ultralytics pkg do all the hard work internally

Release v10.0.24

• Added extra_require for: yolo, evolve and export
• Added error message and suggestions for fixing when boxmot fails to infer inference mode from yolo model name
• Delete timm from requirements as it is only used for CLIP weight initialization and we are not training CLIP-ReID here