Reproduce “Global Features are All You Need for Image Retrieval and Reranking” 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
Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andre Araujo, Bingyi Cao
SuperGlobal, an image retrieval system developed by Peking University and Google Research, exclusively uses global features for both initial retrieval and reranking. This approach achieves state-of-the-art accuracy on benchmarks such as ROxford+1M (71.1% mAP with RN101) and RPar+1M (78.4% mAP with RN101), while offering significant efficiency improvements (64,865x faster and 170x less memory) compared to traditional local feature-based reranking methods.
Key insights
- The combination of GeM pooling with margin-based losses often leads to sub-optimal GeM parameter convergence, which can be improved through post-training tuning.
- Global features can be enriched with regional and multi-scale information through specialized pooling techniques (Regional-GeM, Scale-GeM) to enhance their discriminative power.
- Efficient global-feature-based reranking, inspired by query expansion, can achieve accuracy competitive with or superior to computationally expensive local-feature-based reranking methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.