Reproduce “Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare” 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
Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Baoliang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai and 2 more
Researchers developed Compare2Score, a No-Reference Image Quality Assessment model that leverages a Large Multimodal Model (LMM) to perform human-like visual quality comparisons. The method effectively translates these discrete qualitative assessments into continuous perceptual quality scores, achieving state-of-the-art performance and superior generalization across diverse image distortion types.
Key insights
- Focusing on relative quality comparisons within datasets allows for the principled and robust integration of diverse IQA datasets, overcoming the challenge of inconsistent absolute quality scales.
- Utilizing the full probability distribution (soft comparison) across all LMM output tokens for comparative levels, rather than just the single most probable token, significantly enhances the accuracy of continuous quality score estimation.
- Strategic selection of anchor images with low variability in human subjective ratings (low MOS standard deviation) is crucial for the efficient and unbiased inference of continuous quality scores.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.