A new benchmark called ScenePercept-Bench has been introduced to address the limitations of existing image perception benchmarks, which often apply uniform criteria across diverse scenarios. This benchmark includes 23,100 images across 21 categories and 105 fine-grained scenes, with a hierarchical annotation protocol covering visual quality, aesthetics, and semantics. Additionally, a training framework named Scenario-Conditioned Perceptual Alignment (SCPA) was developed to handle varying perceptual criteria across different scenes, demonstrating state-of-the-art performance on user-generated content image perception tasks. AI
IMPACT This benchmark could improve AI models' ability to understand and evaluate the quality and aesthetics of diverse, real-world user-generated images.
RANK_REASON The cluster describes a new academic benchmark and associated training framework presented in a research paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
