ByteTrack: Multi-object tracking by detection association
Multi-object tracker associating low-confidence detections across frames.
Learn more about ByteTrack
ByteTrack is a multi-object tracking framework implemented in PyTorch that operates on detection outputs from object detectors. The core mechanism associates detection boxes across video frames by leveraging similarity metrics between detections and existing tracklets, including low-confidence detections that would typically be discarded. This approach recovers objects with partial occlusion while filtering background false positives. The system achieves 80.3 MOTA and 77.3 IDF1 on MOT17 at approximately 30 FPS on a single V100 GPU, and has been integrated into multiple detection frameworks including YOLOX.
Low-confidence detection recovery
Unlike conventional trackers that discard detections below a confidence threshold, ByteTrack associates low-score boxes with existing tracklets to recover occluded objects and reduce fragmented trajectories. This approach applies consistently across different detector architectures.
Generic association method
The tracking algorithm operates independently of the underlying detector, allowing it to be applied to multiple state-of-the-art trackers with consistent improvements in IDF1 scores. The method has been validated across different detector implementations.
Real-time performance with accuracy
The implementation balances computational efficiency with tracking accuracy, achieving competitive benchmark results on MOT17 and MOT20 datasets while maintaining frame rates suitable for deployment on single GPU hardware.
Top in AI & ML
Related Repositories
Discover similar tools and frameworks used by developers
LLaMA-Factory
Parameter-efficient fine-tuning framework for 100+ LLMs.
Real-ESRGAN
PyTorch framework for blind super-resolution using GANs.
Wan2.1
Diffusion transformer models for text and image-to-video generation.
ADK
Modular Python framework for building production AI agents.
Awesome Nano Banana
Curated collection of images and prompts from Google's Gemini-2.5-Flash-Image model with model comparisons.