Tracker image input size vs. Detector image size

[adrian-soch]

Here Yolo takes a resized/preprocessed image im:
https://github.com/mikel-brostrom/Yolov5_StrongSORT_OSNet/blob/eab08d1a7bf4b3c63251e9400fe8fe6df6a1c977/track.py#L137

Here the tracker takes the original image im0:
https://github.com/mikel-brostrom/Yolov5_StrongSORT_OSNet/blob/eab08d1a7bf4b3c63251e9400fe8fe6df6a1c977/track.py#L188

Why is this the case?

[mikel-brostrom]

The only reason is to start with high resolution inputs for the ReID model which then resizes to 128x256. I will run an experiment to determine the metric difference on MOT17 between starting with full resolution and 1280.

[mikel-brostrom]

Here are the results: https://github.com/mikel-brostrom/Yolov5_StrongSORT_OSNet/wiki/Effect-of-passing-full-image-input-vs-1280-re-scaled-to-StrongSORT-on-MOT17

	1280	Full images
IDF1	70.199	72.399
MOTA	63.701	63.781
HOTA	55.69	56.783

IDF1 takes the largest hit by far, which makes sense. This behavior would be further accentuated by lowering the resolution even more in val.py. The execution time is barely affected by using the full image. This is because the detections are resized to 128x256 by the ReID preprocessing before they are fed to the ReID model.