Issues related to track IDs

[zhoutianxiang-coder]

Which part of the code matches the object ID between the previous frame and the current frame

[zhoutianxiang-coder]

Does the class and related loss calculation in the code represent different IDs or different classifications?

[zhoutianxiang-coder]

What are the differences between the ID matching process of this model and algorithms like sort?

[zhoutianxiang-coder]

How does the output of the target association process network include tracks_ ID information

[timmeinhardt]

1. At inference time, we do not match. The track queries transfer the identities implicitly. During training, the matching happens here: https://github.com/timmeinhardt/trackformer/blob/e468bf156b029869f6de1be358bc11cd1f517f3c/src/trackformer/models/detr_tracking.py#L81
2. The class loss represents different classes. IDs are not explicitly enforced by the loss.
3. During inference, sort matches IDs based on some costs. We do not match IDs during inference time.
4. Our network does not output track ids.

Your questions indicate a misunderstanding of how TrackFormer generates tracks. We do not perform any matching at inference time. The track queries re-detect objects and thereby transfer identities from frame to frame. Maybe have another look at the related sections in the paper.

[zhoutianxiang-coder]

Where does the information of trackID come from?

[zhoutianxiang-coder]

Track_ Where is the initialization of ID information reflected in the code

[zhoutianxiang-coder]

Whenever a new detection object appears, how to assign an ID number to the new object, and what part of the code does it correspond to?

[zhoutianxiang-coder]

In the paper, it is mentioned that the correlation of trajectories occurs in attention, how to correlate the trajectories of the front and back frames, and where is the specific code embodied?