Reproduce “Compositional Caching for Training-free Open-vocabulary Attribute 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
Marco Garosi, Alessandro Conti, Gaowen Liu, Elisa Ricci, Massimiliano Mancini
COMPOSITIONAL CACHING (COMCA), a training-free method from Marco Garosi and colleagues, enhances pre-trained Vision-Language Models for open-vocabulary attribute detection by employing a sophisticated caching mechanism that accounts for the compositional nature of attributes. The approach outperforms zero-shot and traditional cache-based baselines and achieves competitive performance against recent training-based methods on benchmarks like OVAD and VAW.
Key insights
- Open-vocabulary attribute detection can be effectively performed without task-specific training by carefully leveraging pre-trained Vision-Language Models and external knowledge bases.
- Incorporating attribute-object compatibility, derived from statistical co-occurrence and semantic knowledge from Large Language Models, is crucial for scalable and efficient cache construction.
- Soft attribute labeling in the cache enables better capture of the compositional nature of attributes compared to using hard labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.