Reproduce “Memory Enhanced Global-Local Aggregation for Video 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
Yihong Chen, Yue Cao, Han Hu, Liwei Wang
This work introduces Memory Enhanced Global-Local Aggregation (MEGA), a framework for video object detection that synergistically integrates global semantic information and local localization cues. MEGA achieved state-of-the-art performance on the ImageNet VID dataset, reaching 82.9% mAP with ResNet-101 and 84.1% mAP with ResNeXt-101, by efficiently extending the temporal aggregation range through a novel Long Range Memory (LRM) module.
Key insights
- Combining both global semantic context and fine-grained local localization cues is crucial for robust video object detection, mirroring human visual perception.
- An efficient recurrent memory mechanism (LRM) can overcome the computational challenges of extending temporal aggregation to truly long ranges by caching and reusing features.
- Separately designed relation modules for location-free global interactions and location-based local interactions enhance the effectiveness of aggregated features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.