The above video outlines the construction of Salient Imagenet. Each soft segmentation mask is simply the Neural Activation Map (NAM) for the corresponding feature.
Go to the ANALYSIS page to learn how to easily use Salient Imagenet to evaluate the sensitivity of pretrained models to spurious features, and train improved models.
Scroll below for more brief descriptions of our annotation and processing procedures.
Obtaining Annotations
We utilize three complentary visualization techniques to interpret the function of a neural feature for a class. These are shown above. Together, crowd workers can determine if a feature is core or spurious for a class.
PROCESSING
Salient Imagenet contains five segmentation masks per image, corresponding to the five annotated features per class. We consolidate the five segmentation masks to a single core mask as demonstrated below.
Core mask consolidation is built in to the dataloader code provided in the ANALYSIS tab.
