search

chengche6230 / ReST

[ICCV 2023] ReST: A Reconfigurable Spatial-Temporal Graph Model for Multi-Camera Multi-Object Tracking
MIT License
130 stars 15 forks source link

Question about non-overlapping scenario #5

Closed dlsrbgg33 closed 7 months ago

dlsrbgg33 commented 7 months ago

Hi, thanks for the great work and provided code. As what title says, I am curious about how I can apply your method to non-overlapping scenarios. Do I have to skip the projection feature for association? According to the paper, you also did experiment on Auditorium where the FOV of cameras are not overlapped. I am wondering the training and inference setups for this scenario. It would be appreciated if you can provide guide on this.

chengche6230 commented 7 months ago

Hi, In our work, we focus more on the overlapping FoV scenarios. The sequence of Auditorium is partially overlapping with 2 cameras in both scenes. As for purely non-overlapping scenarios, you need to redesign the feature you need, e.g., rely more on appearance features rather than geometry positions, which have no effect for associating across different views.

As for the experiment details about Auditorium, the projection feature still works. Project each detection in the specific ground plane with appropriate offset. For example, people in the lobby will be projected onto the plane A, which x in [0, MAX_X] and y in [0, MAX_Y], while people in the room will be projected onto the plane B, which x in [0 + OFFSET_X, MAX_X + OFFSET_X] and y in [0 + OFFSET_Y, MAX_Y + OFFSET_Y].

dlsrbgg33 commented 7 months ago

Thank you, it is clear to me.