Reproduce “Multi Positive Contrastive Learning with Pose-Consistent Generated Images” 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
Sho Inayoshi, Aji Resindra Widya, Satoshi Ozaki, Junji Otsuka, Takeshi Ohashi
GenPoCCL uses pose-consistent generated images for contrastive learning in human-centric perception tasks, enabling models to learn robust human structure and appearance representations. The approach achieves improved performance across several benchmarks while often requiring less than 1% of the training data compared to prior methods.
Key insights
- Leveraging pose-consistent image generation creates diverse visual data while maintaining identical human poses, enabling learning of structural features.
- Treating images with the same underlying pose as multi-positive pairs for contrastive learning improves the model's ability to extract pose-invariant features despite appearance variations.
- Incorporating a dedicated [POSE] token within the model architecture specifically enhances the capture of human pose-related features, which is crucial for structural understanding of the human body.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.