Reproduce “MEX: Memory-efficient Approach to Referring Multi-Object Tracking” 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
Huu-Thien Tran, Phuoc-Sang Pham, Thai-Son Tran, Khoa Luu
The MEX method introduces a memory-efficient approach to Referring Multi-Object Tracking (RMOT) by optimizing the cross-modal attention mechanism. It achieves approximately 0.5x lower memory consumption and 1.5x faster inference speeds compared to baseline methods like iKUN, while maintaining tracking accuracy on the Refer-KITTI dataset.
Key insights
- Cross-modal attention mechanisms in Referring Multi-Object Tracking (RMOT) systems, particularly in modular approaches like iKUN, represent a significant memory bottleneck.
- A simplified, custom attention mechanism (MEXAttn) can effectively fuse global, local, and prompt features from different modalities with significantly reduced memory overhead.
- A modular, 'plug-and-play' design allows for integration with existing off-the-shelf trackers without retraining, enhancing their capabilities with referring functionality efficiently.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.