Skip to content
Conference Open access

Data-Efficient Pedestrian Multi-Object Tracking via Zero-Shot Detection and Lightweight DINOv3 Re-Identification

2026 · Proceedings of the 23rd International Conference on Informatics in Control, Automation and Robotics · 0 citations · 38 references

Abstract

: Pedestrian Multi-Object Tracking (MOT) remains challenging in crowded video because missed detections, camera motion, occlusion, and identity ambiguity jointly affect online association. Most high-performing MOT systems rely on pedestrian-tuned detectors and supervised Re-Identification (ReID) models, which increases annotation requirements and obscures whether detection, motion, association, or appearance is the dominant bottleneck. We investigate a data-efficient alternative based on zero-shot detection and lightweight self-supervised appearance adaptation. We combine a D-FINE detector pretrained on Objects365 and COCO, without pedestrian-specific fine-tuning, with DINOv3-based ReID in a modular tracking-by-detection framework. The tracker integrates Kalman prediction, affine Camera Motion Compensation (CMC), cascaded Hun-garian association, and track lifecycle management, with each component separately measurable. On the MOT17 half-validation split, controlled ablations show that detection quality sets the main performance ceiling: replacing zero-shot D-FINE with a MOT17-finetuned YOLOX-X raises Higher Order Tracking Accuracy (HOTA) from 47.2 to 69.9 under the same tracker with raw DINOv3 features. Appearance cues are useful only when identity-discriminative: raw DINOv3 can degrade association, whereas a 99K-parameter attention head on frozen DINOv3 tokens recovers most of the supervised ReID gap. The results indicate that data-efficient pedestrian MOT should prioritize robust detection and use adapted appearance features as a complement to motion and geometry.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.