Deep Self-Taught Learning for Weakly Supervised Object Localization

Abstract

Most existing weakly supervised localization (WSL) approaches learn detectors by finding positive bounding boxes based on features learned with image-level supervision. However, those features do not contain spatial location related information and usually provide poor-quality positive samples for training a detector. To overcome this issue, we propose a deep self-taught learning approach, which makes the detector learn the object-level features reliable for acquiring tight positive samples and afterwards re-train itself based on them. Consequently, the detector progressively improves its detection ability and localizes more informative positive samples. To implement such self-taught learning, we propose a seed sample acquisition method via image-to-object transferring and dense subgraph discovery to find reliable positive samples for initializing the detector. An online supportive sample harvesting scheme is further proposed to dynamically select the most confident tight positive samples and train the detector in a mutual boosting way. To prevent the detector from being trapped in poor optima due to overfitting, we propose a new relative improvement of predicted CNN scores for guiding the self-taught learning process. Extensive experiments on PASCAL 2007 and 2012 show that our approach outperforms the state-ofthe-arts, strongly validating its effectiveness.

Weakly-supervised object detection via mining pseudo ground truth bounding-boxes

“…Most existing methods formulate weakly-supervised detection as an MIL problem [13][14][15][16][18][19][20][21][22] and these approaches divide training images into positive and negative parts, where each image is considered as a bag of candidate object instances. Positive images are assumed to contain at least one object instance of a certain class, and the negative images do not include object instances from this class.…”

“…Some works [11,14,19] try to find better initialization solutions to the above problems, and achieve the gratifying results. For instance, Jie et.…”

“…For instance, Jie et. al [19] propose a self-taught learning approach to acquire tight positive samples by making the detector learn the reliable object-level features, and then re-train itself based on the selected positive samples. The result is that the detector progressively improves the detection ability.…”

“…Besides, many effort s [9,17,19] have been made to improve the optimization strategy for solving the non-convex problem. In [19] , relative improvement of output CNN scores instead of the static absolute CNN score is used at training iterations.…”

“…In [19] , relative improvement of output CNN scores instead of the static absolute CNN score is used at training iterations. These relative improvement scores can effectively filter the suspicious samples whose high predicted scores are from undesired overfitting.…”