An extensive VQA databaseDeep learning approaches have had limited success on existing VQA datasets, either artificial or authentically distorted. We introduce KonViD-150k, an in-the-wild VQA dataset that is substantially larger and diverse, allowing the exploration of training DNNs on massive video collections with coarse annotations.
The database consists of two parts:
KonViD-150k provides a good testing ground for efficient VQA approaches, that are suitable to learn from large collections of videos, and can generalize well based on coarse annotations. Additionally, it is a great tool to investigate VQA methods with different annotation budget distribution strategies. |
Cite usKonViD-150k is freely available to the research community. If you use our database in your research, please cite both references:
@misc{konvid150k,
title = {The Konstanz 150k in-the-Wild Video Database (KonVid-150k)},
author = {G\"otz-Hahn, Franz and Hosu, Vlad and Lin, Hanhe and Saupe, Dietmar},
year = {2021},
url = {http://database.mmsp-kn.de}}
@inproceedings{hahn2021,
title = {KonVid-150k: A Dataset for No-Reference Video Quality Assessment of Videos in-the-Wild},
author = {G\"otz-Hahn, Franz and Hosu, Vlad and Lin, Hanhe and Saupe, Dietmar},
booktitle = {IEEE Access 9},
pages = {72139-72160},
year = {2021},
organization = {IEEE}}
|
KonVid-150k-A setKonVid-150k-B set |
The aggregated scores files
The raw votes files
|