Researchers have developed HiVE-VQA, a novel framework for no-reference video quality assessment (NR-VQA) that utilizes a frozen V-JEPA 2.1 video foundation model. This approach extracts hierarchical features from multiple depths of the pre-trained encoder and fuses them using a compact 3D convolutional head. The HiVE-VQA framework trains only the head, which has significantly fewer parameters than existing complex pipelines, and its inference cost is independent of video resolution. When trained exclusively on LSVQ and evaluated without fine-tuning, HiVE-VQA surpassed 17 state-of-the-art methods on both intra-dataset and cross-dataset benchmarks, demonstrating the efficacy of hierarchical feature fusion from a single frozen model. AI
IMPACT This research demonstrates a more efficient approach to video quality assessment by leveraging frozen foundation models, potentially reducing computational costs and improving performance in video analysis tasks.
RANK_REASON The item describes a new research paper detailing a novel framework for video quality assessment. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
