There is 40-50% drop in precision with state of the art results if test images are ill-formed. The imagenet dataset used is far from ready for real world use cases. A bunch of IBM and MIT researchers are trying to fix this - https://objectnet.dev/