Quantization vs. Identity Preservation in Real-Time MOT
Overview
This project evaluates how quantization — reducing model precision to INT8 or FP16 via TensorRT — affects identity-tracking quality in real-time multi-object tracking on embedded hardware.
The problem
Quantization speeds up inference, but it introduces bounding-box jitter. In the association stage of trackers like ByteTrack and BoT-SORT, that jitter causes ID switches — so identity metrics (HOTA, IDF1, MOTA) degrade disproportionately compared to raw detection accuracy (mAP).
Setup
- Hardware: NVIDIA Jetson Orin Nano vs. x86 baseline
- Precision modes: FP32, FP16, INT8 (TensorRT)
- Metrics: HOTA, IDF1, MOTA, detection mAP, FPS
Results
(Add your charts, tables, and GIFs here once you have them.)
To add an image, drop the file into the public/ folder and reference it:
