search

selflein / GraphNN-Multi-Object-Tracking

Unofficial PyTorch implementation of "Learning a Neural Solver for Multiple Object Tracking"
122 stars 23 forks source link

How to Modify for Multiple Target Tracking #5

Open bayraktare opened 4 years ago

bayraktare commented 4 years ago

Hi, thanks for sharing this valuable code.

I just want to learn if we can modify your approach to train on our custom datasets, which may contain multiple classes other than pedestrian only? If yes, can you explain a little bit how?

Thanks.

selflein commented 4 years ago

Hi, thank you for your interest.

First, I want to mention that I am not the author of the paper and just re-implemented the approach (for the most part). I guess the simplest extension for mutiple object classes would be to create a separate graph for detections of each object class and run the algorithm independently. I could imagine that the encoders for the features of the edges and the classifier of the edges might generalize across object classes so one can share these parameters.

What is remaining would be to train a re-identification model on all object classes OR a re-identification model for each object class independently which provide the node embeddings. One might need to experiment what works better/ there might also be literature on that.

I hope this helps. I guess there are more sophisticated ways to deal with multiple object classes, e.g., creating one graph with all detections and adding the object class as a feature to the nodes. I could imagine that also working well (but people/objects seldomly change object class so might just create additional overhead).

If you proceed to experiment with this, I would be really curious about your results.

qintaohu commented 4 years ago

Hi, I modified the detector to detect a specific category of people. Can I directly replace the weight of fastercnn. Thanks!