fix: 🐞 allow None as a valid model_or_checkpoint_path in DeepSORTFeatureExtractor

[onuralpszr]

Fixed errors are

trackers/core/deepsort/feature_extractor.py:116: error: Item "None" of "Any | None" has no attribute "to"  [union-attr]
trackers/core/deepsort/feature_extractor.py:117: error: Item "None" of "Any | None" has no attribute "eval"  [union-attr]
trackers/core/deepsort/feature_extractor.py:154: error: "None" not callable  [misc]
trackers/core/deepsort/tracker.py:170: error: Missing positional argument "model_or_checkpoint_path" in call to "DeepSORTFeatureExtractor"  [call-arg]
Found 4 errors in 2 files (checked 14 source files)

[soumik12345]

IMO, having allowing a None value only makes sense if we can set a default model instead of raising a ValueError.

[onuralpszr]

@soumik12345 which model ? :) I didn't wanna slap resnet50 or something please let me know which :)

[soumik12345]

Ideally, a reproduction of the original feature extractor architecture described in Section 2.4 and Table 1 of the DeepSORT paper should be treated as a default model (downloaded from a default URL). However, since the paper doesn't reveal the training methodology of the model, we will need to research a little bit regarding what might be the best way to reproduce the results from the paper.

[soumik12345]

@SkalskiP I would like to know your thoughts on this as well.

[rolson24]

@onuralpszr Not sure that would follow the clean-room implementation goals of this project, as the DeepSORT code is not open license.

Alternatively, we could also use models from Fast-Reid which is Apache 2.0. The models from Fast-ReId are still being used in SOTA papers. For example DeepOC-SORT uses SBS50 from Fast-ReId:
"Implementations Our implementation is based on OCSORT [2, 28]. We use the same YOLOX detector as recent
works [4, 3, 13, 1] to make a fair comparison of tracking performance. For Re-ID, we use SBS50 from the fast-reid [29] library." (DeepOC-SORT paper)

The DeepOC-SORT paper appears to use the ReID model to extract the features for each detection, similar to our implementation of DeepSORT:
"In previous work [3, 4], the deep visual embedding used to describe a tracklet is given by an Exponential Moving Average (EMA) of the deep detection embeddings frame by frame." (DeepOC-SORT paper)

Alternatively, we could reproduce the DeepSORT architecture described in the paper, and train on Market-1501, which doesn't seem to have a license (couldn't find one?) and is widely available on Kaggle.

Other ReID datasets have academic only licenses, but it seems like some ReID models trained on them are open-license (probably want to avoid this)? Some datasets have also been retracted due to privacy issues

[SkalskiP]

I agree with @soumik12345's point here:

allowing a None value only makes sense if we can set a default model instead of raising a ValueError.

I'll be committing my related updates to this PR shortly. As @soumik12345 noted, while our long-term goal is to publish our own fine-tuned weights, the current default pathway involves initializing the feature extractor with a timm model, like so:

from trackers import DeepSORTTracker, DeepSORTFeatureExtractor

feature_extractor = DeepSORTFeatureExtractor.from_timm("resnet50d.ra2_in1k")
tracker = DeepSORTTracker(feature_extractor)

While this initialization is slightly verbose, we intend to simplify it ahead of the v2.0.0 release.

[CLAassistant]

All committers have signed the CLA.