Datasets for the evaluation of explainability methods for computer vision models
Set of datasets of images, their ground truth, their saliency map for one model, and the manual annotations of the saliency maps with semantic concepts of various granularities. The images in each set were selected in order to inject class-specific biases in the dataset.
- One set consists of two classes sampled from ImageNet, representing ambulances and vans.
- One set consists of two classes sampled from PA-100K, representing human males and females.
- 4 sets consists of three classes extracted from ImageNet, lobsters, sharks, and trouts, sampled in different ways to include various biases.
- 2023-03-30 first online, published, posted