A tendency for object decisions to follow local texture rather than global shape under cue conflict. Human–model comparisons on matched conflict images can reveal different strategies despite similar ordinary classification accuracy. Altered training distributions can change the bias without changing the network architecture.[ref]
ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness by Robert Geirhos and others. Abstract and cue-conflict experiments; ICLR 2019, texture–shape conflict and Stylized-ImageNet training
Articles by others on the same topic
There are currently no matching articles.