Feedback IQA dataset

For this study, we used a dataset of five source images that were compressed by two codecs, VVC Intra and JPEG XL from JPEG AIC-3 database. This resulted in 10 sequences of 11 cropped images each, starting with the source at distortion level 0 and increasing levels from 1 to 10.
​
We conducted our experiment on the MTurk platform. We posted one human intelligence task (HIT) with 200 assignments for 200 unique crowdworkers. In each assignment, besides the training questions, a worker had to answer 120 study and trap questions in random order, once with feedback and then again in different random order without feedback.

本研究中，我们采用了由两种编解码器——VVC Intra及JPEG XL（源自JPEG AIC-3数据库）压缩的五张原始图像所构成的数据集。由此生成了包含11张裁剪图像的10个序列，起始图像的失真级别为0，随后逐级提升至10级。实验在MTurk平台上进行，我们发布了一项包含200个任务的人工智能任务（HIT），分配给200位不同的众包工作者。在每一项任务中，除了训练问题之外，工作者需要按照随机顺序回答120个研究问题和陷阱问题，首先在提供反馈的情况下进行，随后在另一种随机顺序且不提供反馈的情况下再次作答。