Training code

[ITBoy-China]

Hello, thanks for your great work! I successfully run the inference code and I'd like to use my own dataset to train it. However, I couldn't find the training code. Have you already released the training code? Thanks again.

[George-Zhuang]

Hi ITBoy,

Thanks for your interest in NetTrack. NetTrack is primarily designed for zero-shot generalization tasks, but you may still fine-tune it on your own data by fine-tuning the base model GroundingDINO. Since the official GroundingDINO has not given the fine-tuning code, it is recommended to use the MMGroundingDINO instead. But it is still a little complicated.

If the fine-tuning work is not urgent, please be patient for our extended work on NetTrack, where more models and fine-tuning will be supported more conveniently. However, if you still want to fine-tune NetTrack now, please continue to raise any difficulties you encounter under this issue.

[mrtpk]

Hi @George-Zhuang,

Thank you for sharing this amazing work.

When you said "zero-shot generalization" does it mean I can use it with other custom-trained object detectors? I use the YOLO object detector and ByteTrack to track underwater objects. I want to try out your algorithm by replacing ByteTrack.

Any pointers or input is greatly appreciated. Thanks.