OBJECT DETECTORS EMERGE IN DEEP SCENE CNNS

CNN works well on diff. datasets and tasks, but what does it learn?

• this paper: CNN learns object detector in an unsupervised manner when trained on scene understanding tasks.
• simplifying images: at each iteration, remove a segment which produces the smallest decrease in classification score
  • key objects remained
• empirical reception fields: smaller(~1/2) than theoretical size
• compared to CNN trained on Imagenet, CNN trained on scene understanding focus more on object in high level layers
  • object detector as internal representation of scene understanding?
• AP: ranging to >90%!