-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
Q: What datasets does FlashTrack support? A: Any dataset in MOT17/MOT20 format (images + gt.txt annotations). Convert your data to this format for custom tracking.
Q: What's the difference between FlashTrack and FlashDet? A: FlashDet is for object detection (finding objects in frames). FlashTrack is for multi-object tracking (maintaining identity across frames). They share the same ShuffleNetV2 backbone.
Q: Can I use FlashTrack without a ReID model? A: Yes! ByteTracker and SORTTracker work without ReID features. Only DeepSORTTracker benefits from ReID.
Q: How much data do I need? A: The MOT17 training split (~5K identities) works well. For custom domains, at least 100+ identities with 10+ samples each.
Q: Which LoRA variant should I use?
A: Start with standard. Use dora for higher quality, lora_fa for minimal memory, or lora_plus for faster convergence.
Q: How do I resume training?
A: Use Trainer(resume="workspace/checkpoint_last.pth") or --resume flag.
Q: How do I export to ONNX?
A: flashtrack export --model best.pth --output reid.onnx --simplify
Q: What's the smallest model? A: FlashTrack-m-0.5x at ~0.3M params / ~0.6 MB FP16.
Q: Can I run on CPU?
A: Yes, use --device cpu. CPU inference is fast enough for real-time with ByteTracker.
Q: scipy import error?
A: Install scipy: pip install scipy>=1.10.0
Q: CUDA out of memory?
A: Reduce batch size, enable amp=True, or use LoRA/QLoRA.
Q: Track IDs keep changing?
A: Increase track_buffer, lower match_thresh, or switch to DeepSORTTracker with ReID features.
FlashTrack — Multi-object tracking | PyPI | MIT License