Reproduce “MV-YOLO: Motion Vector-aided Tracking by Semantic Object Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Saeed Ranjbar Alvar, Ivan V. Bajić
MV-YOLO introduces a two-stage hybrid visual object tracker that leverages motion vectors from HEVC bitstreams to generate an approximate region of interest, which is then refined by a semantic object detector (YOLO) on fully decoded frames. The framework achieves superior accuracy (best average Success AUC with YOLOv3) and improved efficiency compared to several existing pixel-domain trackers.
Key insights
- Efficient reuse of motion vectors (MVs) directly from the compressed video bitstream provides a computationally inexpensive way to obtain an initial, approximate target location.
- Combining this motion-based approximation with a robust, general-purpose semantic object detector enables pixel-precise localization and inherent robustness to object appearance, scale changes, and viewpoint variations.
- An adaptive IOU threshold for selecting the final bounding box allows the tracker to maintain target association during temporary occlusions or challenging detection conditions, improving overall robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.