Reproduce “Query2Label: A Simple Transformer Way to Multi-Label Classification” 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
Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, Jun Zhu
The Query2Label (Q2L) framework introduces a simple Transformer-based method for multi-label image classification that leverages learnable label embeddings as queries within a Transformer decoder. This approach adaptively extracts class-specific features from image regions, achieving state-of-the-art performance on various benchmarks including 91.3% mAP on MS-COCO and 96.6% mAP on PASCAL VOC 2012.
Key insights
- Transformer decoders, particularly their cross-attention mechanism, are effective for adaptively extracting local discriminative features pertinent to specific labels in multi-label classification.
- End-to-end learning of label embeddings allows the model to capture inter-label relationships implicitly directly from the data, simplifying the architecture compared to relying on pre-computed correlations.
- Multi-head attention can focus on distinct parts of an object, contributing to enhanced robustness against occlusion and viewpoint changes, while also providing interpretability into model decisions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.