Mike

Thank you for your PR! So many Yolo versions popped up lately that I don't know what to make out of them... I need to figure out if I want...

I am using a Yolov5 model to generate detections which I pass to both StrongSORT and BoTSORT. I get worse results on MOT16 using BoTSORT with default parameters than StrongSORT...

I identified the slow part. step2 in `bot_sort.py`: https://github.com/NirAharon/BoT-SORT/blob/03a3a6851013abd54b95d863ec765c1437a63de8/tracker/bot_sort.py#L291 Some example of timed steps: ```bash 0: 480x640 1 person, Done. YOLO:(0.045s), BoTSORT:(0.092s) step2: 0.46398663699999965 step3: 0.0002665390000000656 step4: 6.4769999994496175e-06 step5: 6.332999999969502e-06...

Ok, the camera motion compensation seems to be the issue: ```python # Fix camera motion warp = self.gmc.apply(img, dets) STrack.multi_gmc(strack_pool, warp) STrack.multi_gmc(unconfirmed, warp) ```

I have the same issue. https://github.com/NirAharon/BoT-SORT/issues/39 But way more extreme in my case. BoTSORT is not something I would use for real-time applications.

> What's your findings for the most performant real time tracking algorithm @mikel-brostrom? Both OCSORT and ByteTrack are absolutely the fastest, achieving around 800FPS on a mere Intel® Core™ i5-4210U...

If you have exported the models to the fastest framework running on your desired hardware I would say there is very little more to do more than changing tracking module....

You could pass both the bboxes and pose information to: https://github.com/mikel-brostrom/Yolov5_StrongSORT_OSNet/blob/4cf9555609ae2c6c9a9dba7f4202e0a509ef90b5/track.py#L209 And store the pose associated to as specific bbox in the track.

Conversion to TFLite can be achieved. It is cumbersome as you will need to do something like: 1. PyTorch -> ONNX (https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html) 2. ONNX -> OpenVino (https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_ONNX.html) Notice that most...

Wow, nice, I am glad that it worked! The current code in StrongSORT does not support TFLite model loading. `torch.load` is not meant for TFLite models. Try this: https://stackoverflow.com/questions/50443411/how-to-load-a-tflite-model-in-script. Consider...