Reproduce “Vision-Language Models Can't See the Obvious” 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
Yasser Dahou, Ngoc Dung Huynh, Phuc H. Le-Khac, Wamiq Reyaz Para, Ankit Singh, Sanath Narayan
The Technology Innovation Institute introduces Saliency Benchmark (SalBench) to assess Large Vision-Language Models (LVLMs) on low-level visual saliency detection. Evaluations reveal that state-of-the-art LVLMs achieve low accuracy on tasks trivial for humans, with the vision backbone identified as a key limitation.
Key insights
- Large Vision-Language Models (LVLMs) exhibit surprising limitations in detecting low-level visually salient features, a task trivial for humans.
- The vision backbone component, rather than the Large Language Model's conceptual understanding of saliency, is the primary bottleneck in LVLM performance on these tasks.
- Standard LVLM architectures and training methodologies appear unsuitable for effectively capturing and integrating low-level saliency information, even with targeted fine-tuning data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.