Short Video Quality Assessment Competition (KVQ) at CVPR NTIRE 2024

The CVPR NTIRE Workshop, a leading computer‑vision conference, partnered with the Intelligent Media Lab at the University of Science and Technology of China and Kuaishou Technology to launch the first Short Video Quality Assessment competition at CVPR NTIRE 2024, aiming to establish quality‑evaluation benchmarks for short videos.

Short videos have become a dominant media format, but variations in creation, processing, and transmission lead to significant quality differences, making subjective quality measurement a critical challenge for platform development.

The competition uses the KVQ dataset, a large‑scale short‑video quality assessment collection comprising 4,200 videos covering nine content scenes (e.g., landscape, crowd, food, portrait). The data are split into 70% training, 10% validation, and 20% test sets; participants receive the training and validation sets and submit results via CodaLab for evaluation.

Key competition resources include the competition website (https://codalab.lisn.upsaclay.fr/competitions/17638), project page (https://lixinustc.github.io/projects/KVQ/), the accepted CVPR 2024 paper (https://arxiv.org/abs/2402.07220), and the CVPR NTIRE 2024 workshop site.

The contest is open to individuals, academia, research institutes, and companies. Teams may have up to eight members, each team can submit only one final result, and any use of external data must be disclosed in the final report.

Prizes are awarded to the top three teams: 1st place receives $1,000 plus a certificate, 2nd place $600 plus a certificate, and 3rd place $400 plus a certificate, provided by Kuaishou or the NTIRE organizers.

Organizers are Kuaishou Audio‑Video Technology Department and the USTC Intelligent Media Lab. Contact persons include Xin Li ([email protected]), Kun Yuan ([email protected]), Yajing Pei ([email protected]), Yiting Lu ([email protected]), Ming Sun ([email protected]), Radu Timofte ([email protected]), Chao Zhou ([email protected]), and Zhibo Chen ([email protected]).

The KVQ dataset was collected from Kuaishou’s platform and annotated by USTC, featuring diverse creation modes (e.g., three‑segment, effects, subtitles, live streams) and typical processing pipelines (enhancement, pre‑processing, transcoding). A baseline algorithm, KSVQE, built on a 3D Swin Transformer with CLIP content priors and region‑adaptive sampling, demonstrates strong performance on the KVQ benchmark.

Additional resources and related links are provided for participants to explore the dataset, baseline code, and competition details.