Answers are grounded in this paper and cite their sources.
Introduction Separate appearance encoders add computational cost to multi-object tracking. This study examines confidence-ranked descriptor retention in a modified DeepSORT pipeline that reuses multi-scale YOLOv5s neck features for vehicle appearance association. Methods Raw predictions are mapped to detector-derived descriptors and clustered using spatial overlap. The parameter L limits retained feature vectors per dimensionality, and appearance costs use minimum cosine distance between compatible descriptors. Experiments on one highway-surveillance sequence compare SORT, DeepSORT, StrongSORT, OC-SORT, ByteTrack, and proposed configurations with L = 5, 10, and 20 under limits of four CPU cores and 4 GB of RAM. Association disagreement is evaluated against a visually reviewed, DeepSORT-initialized identity reference. Evaluation uses one random pattern per nonzero dropout rate, scores only frames with available detections, and suspends track deletion during dropout. Results Mean tracker-side processing times are 6.83–11.38 ms per frame, representing 15.2–25.3% of DeepSORT’s measured time; detector inference and video decoding are excluded. For L = 20, association disagreement is 0.43%, 0.96%, and 1.69% at frame-level detection-dropout rates of 0%, 25%, and 50%, respectively. DeepSORT and StrongSORT achieve lower disagreement. Discussion These findings describe a trade-off between processing cost and association agreement under the tested protocol. Standard tracking metrics, public-benchmark generalization, and end-to-end real-time performance remain unevaluated.
No takes yet. Share an insight, caveat, or question.
El-Alami et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: