Reproduce “More Grounded Image Captioning by Distilling Image-Text Matching Model” 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
Yuanen Zhou, Meng Wang, Daqing Liu, Zhenzhen Hu, Hanwang Zhang
A novel framework improves visual grounding accuracy in image captioning by distilling knowledge from a Part-of-Speech-filtered image-text matching model (POS-SCAN) and introducing a combined reward for Self-Critical Sequence Training. The method enhances attention metrics, such as increasing F1_all from 5.11% to 6.47% and F1_loc from 14.67% to 17.96% on Flickr30k, without requiring explicit region-word annotations.
Key insights
- Standard image-text matching models, like original SCAN, may not inherently possess superior visual grounding compared to baseline captioning models.
- Filtering non-visual words (e.g., using Part-of-Speech tags for nouns) in image-text matching models significantly enhances their intrinsic grounding accuracy, making them effective teachers.
- Optimizing image captioners for standard sequence-level metrics like CIDEr via SCST can negatively impact visual grounding performance, highlighting a conflict between these objectives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.