CompressedVQA-AEV: Full-Reference and No-Reference Quality Assessment Models for Asymmetric Encoded Videos

2026-07-06 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

This report details the CompressedVQA-AEV-FR and CompressedVQA-AEV-NR models, developed as solutions for the QoMEX 2026 Grand Challenge on Video Quality Assessment for Asymmetric Encoded Videos. The full-reference (FR) model, CompressedVQA-AEV-FR, utilizes a Swin-B backbone to extract multi-stage similarity statistics between reference and distorted videos for quality prediction. For the no-reference (NR) setting, CompressedVQA-AEV-NR employs complementary frame-level encoders based on SigLIP2 and Swin-B, followed by temporal mean pooling and cross-fold ensembling to estimate perceptual quality. CompressedVQA-AEV-FR secured first place in its track, while CompressedVQA-AEV-NR achieved fourth place in the NR track, demonstrating their effectiveness. The code for these models is publicly available on GitHub.

Key takeaway

For video quality assessment engineers or researchers focused on asymmetric encoded videos, you should evaluate the CompressedVQA-AEV models. CompressedVQA-AEV-FR achieved first place in the QoMEX 2026 Grand Challenge FR track, indicating its strong performance for full-reference scenarios. Consider integrating these models, particularly the FR solution, into your assessment pipelines or research to leverage their demonstrated effectiveness. The publicly available code facilitates adoption and experimentation.

Key insights

CompressedVQA-AEV-FR and -NR models provide effective full-reference and no-reference video quality assessment for asymmetric encoded videos.

Method

The FR model uses a Swin-B backbone for multi-stage similarity statistics. The NR model combines SigLIP2 and Swin-B frame encoders with temporal mean pooling and cross-fold ensembling.

Topics

• Video Quality Assessment
• Full-Reference VQA
• No-Reference VQA
• Asymmetric Video Encoding
• Swin-B
• SigLIP2

Code references

• sunwei925/CompressedVQA-AEV

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Learning Perceptual Representations for Gaming NR-VQA with Multi-Task FR Signals — MTL-VQA uses multi-task learning with FR signals to enable label-efficient NR-VQA for gaming videos.
• AdaCodec: A Predictive Visual Code for Video MLLMs — AdaCodec uses a predictive visual code to reduce redundancy in video MLLMs, improving efficiency and performance.
• VMAF v1: Good Is Not Good Enough — VMAF v1 improves video quality assessment by integrating advanced perceptual features and a unified model for diverse viewing conditions.
• ViQ: Text-Aligned Visual Quantized Representations at Any Resolution — ViQ balances semantic richness and detail preservation in discrete visual representations for efficient, unified multimodal modeling.
• Analytic Score Optimization for Multi Dimension Video Quality Assessment — Multi-dimensional VQA with interpretable annotations and a novel optimization method improves video quality assessment.
• CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long-Context LLM Inference — CompressKV uses Semantic Retrieval Heads and layer-wise budget allocation to efficiently compress KV caches for long-context LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.