Datasets and benchmarks released by our group, newest first. Each card links to the data and, where there is one, to the paper to cite.
A benchmark for human-object interaction editing, released with InteractEdit. Each source image comes with subject, object and background crops and masks, plus target interactions to edit towards. IEBench-L extends it to 100 source images and 326 pairs.
Fair Evaluation for Affective Recognition: standardized, valence-balanced seen and unseen class splits of FI, Emotion6 and ArtPhoto for zero-shot emotion recognition, scored with the FaH metric. The images themselves come from those original datasets.
Integrated-circuit text spotting and aesthetic assessment on real-world images, with aesthetic classes annotated for every character.
Bilingual scene-text visual question answering with an evidence-based measure: a model must also give a bounding box for its predicted answer. The repository hosts the images, the train and test annotations and an evaluation script.
Our WikiArt split, for replicating or fairly comparing with our fine-art classification and ArtGAN papers. All images come from WikiArt.org; we are not responsible for their content or meaning.
