AVT-VQDB-UHD-1-NVC - A Subjective and Objective Video Quality Dataset for Neural and Traditional Video Codecs

This comprehensive dataset presents subjective and objective video quality evaluation for both neural and traditional video codecs using 4K/UHD-1 resolution videos. The research, to be presented at Picture Coding Symposium (PCS) 2025 in Aachen, Germany, addresses the critical question of whether existing quality metrics remain valid for evaluating neural video codecs (NVCs) as they emerge as promising alternatives to traditional compression methods.

Author: Benjamin Herb, Rakesh Rao Ramachandra Rao, Steve Göring (Technische UniversitÀt Ilmenau), Alexander Raake (RWTH Aachen University)

Partner: No

Tags: , , , , , , , , ,

Categories:

DOI: 10.48550/arXiv.2511.00969

Subjective scores: true

Total: 216

SRC: 6

Ratings: 6480

Resolution: 360p, 720p, 1080p, 2160p

Method: 5-point Absolute Category Rating (ACR)

Description

This comprehensive dataset presents subjective and objective video quality evaluation for both neural and traditional video codecs using 4K/UHD-1 resolution videos. The research, to be presented at Picture Coding Symposium (PCS) 2025 in Aachen, Germany, addresses the critical question of whether existing quality metrics remain valid for evaluating neural video codecs (NVCs) as they emerge as promising alternatives to traditional compression methods.

The dataset contains 216 encoded sequences derived from 6 source videos (8-10 seconds each, 4K/UHD-1, 60 fps, YUV420p 8-bit format) selected from the AVT-VQDB-UHD-1 database. Videos were encoded at four resolutions (360p to 2160p) using nine different quality parameters with four codecs: traditional codecs AV1 (v3.12.0) and VVC/H.266 (vvencFFapp v1.13.1), and neural codecs DCVC-FM and DCVC-RT.

Subjective testing involved 30 participants (students and employees) with 26 passing outlier detection (Pearson correlation >0.75), collecting 6,480 ratings (33 excluded due to technical errors) using the 5-point Absolute Category Rating (ACR) method. The controlled viewing environment used a 43-inch UHD monitor at 1.5H viewing distance.

The research evaluated 13 quality metrics across full-reference (PSNR, SSIM, MS-SSIM, VMAF variants, CVQA-FR, LPIPS), no-reference (MUSIQ, CVQA-NR, FasterVQA, Dover, Q-Align), and hybrid (AVQBits H0 f) categories. Results showed that VMAF and AVQBits H0 f demonstrated strong Pearson correlation, while FasterVQA excelled among no-reference metrics. Notably, metric reliability showed no significant performance differences between traditional and neural video codecs, validating the use of existing quality metrics for evaluating neural codec performance.

Access

Openly available for download from https://avtshare01.rz.tu-ilmenau.de/avt-vqdb-uhd-1-nvc/ (source videos ~13 GB, processed videos ~74 GB)

License

Open access