Zero-Shot YOLOv8 Descriptors for Low-Latency Object Tracking

Object tracking in real-time applications often relies on the tracking-by-detection paradigm, where object localization and temporal association are decoupled. While methods such as DeepSORT integrate a detector like YOLO with additional CNN-based ReID module for appearance matching, the ReID network imposes significant computational and memory burdens, especially on edge or CPU-only systems. This work explores the use of zero-training appearance descriptors extracted directly from YOLOv8's internal multi-scale feature maps (P3, P4, P5), used individually as lightweight alternatives to conventional ReID embeddings. Our method selects a feature scale based on object size and retrieves a fixed-length feature vector via a simple center-cell lookup. On the MOT20 train benchmark, our fastest configuration, (i.e. based on the P3 feature map) achieves more than 90% of the HOTA score of the standard DeepSORT with ReID, while reducing tracking time per frame from 0.279 s to 0.035 s on a CPU. These results demonstrate that YOLOv8's native features hold sufficient discriminative power for training-free, real-time object association, eliminating the need for costly ReID networks in resource-constrained deployments.