Is there a way to have your own detection module before passing it to the tracking? or a TensorRT optimisation for the existing checkpoints?

[R-C101]

Basically I want to be able to customise the detections so that I can choose what classes need to be tracked or even better if I can have my own pre-trained detection models like RF DETR/YOLO for use with this.

[HELLORPG]

I'm sorry , I'm a bit unclear about the second half of your question.

If you want to filter out certain detections using custom rules, you can add the corresponding code between the Detector and the ID Decoder. For example, I suggest you customize the specific rules of the function _get_activate_detections at this line to select the desired detection results for tracking.

MOTIP/models/runtime_tracker.py

Line 94 in e23387c

scores, categories, boxes, output_embeds = self._get_activate_detections(detr_out=detr_out)

[R-C101]

Thanks for the reply, I was asking if there could be a modification to the code such that any Object Detection model can be coupled with the MOTIP tracker and not just DETR, for example Ultralytics YOLO and RFDETR from roboflow

[HELLORPG]

Theoretically, YES. You just need to ensure that the ID Decoder can receive the object features and the bounding boxes.

Replacing DETR-family detectors may be straightforward, as most DETR-series code is derived from the original DETR implementation. However, replacing traditional detectors, such as those from the YOLO series, would involve more complex code modification. That's why we chose to use DETR-based detectors: we have more experience with them.