Steering Self-Supervised Feature Learning Beyond Local Pixel Statistics

摘要: We introduce a novel principle for self-supervised feature learning based on the discrimination of specific transformations an image. argue that generalization capability learned features depends what image neighborhood size is sufficient to discriminate different transformations: The larger required and more global statistics can describe. An accurate description allows better represent shape configuration objects their context, which ultimately generalizes new tasks such as object classification detection. This suggests criterion choose design transformations. Based this criterion, we transformation call limited context inpainting (LCI). inpaints patch conditioned only small rectangular pixel boundary (the context). Because information, inpainter learn match local statistics, but unlikely claim same be used justify performance rotations warping. Indeed, demonstrate experimentally LCI, warping rotations, yields with state art capabilities several datasets Pascal VOC, STL-10, CelebA, ImageNet. Remarkably, our trained achieve Places par through supervised ImageNet labels.