AUTOMATICALLY LOCALIZING THE RELEVANT IMAGE REGIONS FOR WEAKLY-SUPERVISED VISUAL RECOGNITION

Our group's recent and ongoing work on weakly-supervised visual recognition will be presented. In contrast to fully-supervised algorithms, the proposed methods do not require detailed localization annotations during training, and instead can learn to attend to the relevant visual regions given only image-level semantic tags that state whether an object is present or absent in the image (e.g. an image tagged with "car") or pairwise comparisons that state whether one image has more of a visual property than the other ("person A is more smiling than person B"). I will show how the proposed algorithms can produce state-of-the-art weakly-supervised results for object detection and attribute modeling, and can sometimes reach accuracy that is (nearly) on par with fully-supervised methods at a fraction of the annotation cost.

Speaker: Yong Jae Lee, UC Davis