CLASP on CPUs: Measured Energy, Deadline Compliance and Exact Computation Skipping for Multi-Person Tracking
Year: 2026
Type: preprint
Abstract: Multi-person tracking on CPU-only platforms is usually reported in frames per second, with energy inferred rather than measured. We measure per-stage latency, per-frame deadline compliance, cores used and Intel RAPL package energy of five tracking-by-detection methods—CLASP, ByteTrack, DeepSORT, the official OC-SORT and TrackTrack—with isolated processes, seeded run orders, idle baselines and hash-verified inputs and outputs, over 76 sequences of MOT17, MOT20 and DanceTrack and three detector scales. We also study exact computational reduction of association: a component-wise dominance certificate decides in linear time when a bounded appearance term cannot change the optimal assignment, and four exact reformulations remove redundant work. The certified CLASP produces byte-identical output on 158 sequences at 0.64× the tracker energy of its reference. We prove that the certificate can never succeed once the appearance weight reaches the acceptance threshold, and measure the consequence: revising CLASP for accuracy raised its tracker energy by 37% (95% CI 22–53%), turning a tracker as cheap as ByteTrack into one 1.4× as expensive. Association uses one core and meets a 30 fps deadline with a margin above 16×. With YOLO11n in the loop, lightweight trackers average 30 fps, yet 36–46% of frames exceed 33.3 ms; camera-motion compensation raises system energy by 43% and removes that operating point. A vectorised cost matrix reduced operations but increased energy. Scheduling the detector over time cost accuracy, and crop verification saved 21–41% on design sequences but failed its pre-registered held-out confirmation.
Source: Research Square
Author: SANTIAGO BAQUERO DELGADO
Institution: University of La Salle
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236412903
