15 Computer Vision News Motivated by these limitations, our paper proposes a synthetic classification dataset that is specifically designed for the partbased evaluation of XAI methods. It consists of renderings of funnylooking birds of various ‘species’ on which ‘semantically meaningful’ image-space interventions can be performed to approximate groundtruth explanations. Following a similar idea as the above feature deletion protocols, the dataset allows to delete individual parts of the birds, e.g., their beak or feet, to measure how the output of the model drops. If a deleted part causes a large drop in the output confidence, one can assume that the part is more important than one that only causes a minor output drop (Fig. 1). This allows to move from the above pixel level to a semantically more meaningful part level and, as the training set now includes images with deleted parts, all interventions can be considered in domain. To thoroughly analyze various aspects of an explanation, the FunnyBirds framework considers three dimensions of XAI quality and two dimensions of model quality (Fig. 2). Various interesting findings were made, using the proposed FunnyBirds framework. First, architectures that were designed to be more interpretable, such as BagNet, often achieve higher metrics than the corresponding standard backbone networks. Second, the VGG16 backbone appears to be more explainable than theResNet-50backbone, FunnyBirds Fig 1. Removing individual bird parts and measuring the output change allows to approximate ground-truth importances for each part. In this example, the beak is more important than the feet. BEST OF ICCV 2023
RkJQdWJsaXNoZXIy NTc3NzU=