LITE and PRIME: Robust Streamlined MOT Solutions for Real-Time Systems

초록

Multi-object tracking (MOT) underpins applications from surveillance to autonomous driving, yet appearance-based identity preservation often requires dedicated re-identification networks that limit real-time deployment. This work extends the Lightweight Integrated Tracking-Feature Extraction (LITE) paradigm by systematically analyzing appearance descriptors derived directly from detector feature maps without requiring a separate model. We formalize the feature extraction process, examine its computational behavior, and evaluate feature representations across multiple detector scales, network layers, and tracking frameworks. To better understand feature quality, we present an evaluation procedure that measures discriminative strength using similarity distributions. Evaluations follow a holistic pipeline assessment covering detection through tracking. Notably, research provides real-world, robustness of MOT systems across unseen tracking environments. This benchmark is achieved owing to proposed evaluation protocol which does not allow any training of either detectors and trackers. Experiments on MOT17 and MOT20 demonstrate that LITE-equipped trackers maintain comparable HOTA scores to their ReID-based counterparts while running 2- 5 x faster, exceeding 50 FPS on a single NVIDIA RTX 3090 GPU (24 GB). These findings show that detector-embedded appearance modeling provides a scalable and deployment-oriented alternative to traditional re-identification pipelines. The source code is available at https://github.com/Jumabek/LITE which provides details steps to reproduce our experiments.

키워드