Convolutional Neural Network Based Mixed Supervised Object Detection

Object detection is a core problem in computer vision and pattern recognition.The labeling cost of large number of bounding boxes is one of the main challenges for training modern object detectors.To reduce the dependence on expensive bounding box annotations,the mixed supervised object detection(MSOD)task is studied in this thesis,in which a small proportion of fully-annotated images(i.e.,bounding box level annotation is given)and a large scale of weakly-annotated images(i.e.,only the image level annotation is given)are used to train an effective detector.An end-to-end detection architecture for MSOD is proposed in this thesis.Different from previous weakly-supervised or mixed supervised object detectors which rely on hand-craft object proposals,the proposed architecture comprises a region proposal network,thus enabling nearly cost-free high-quality region proposals.Besides,it can take both fully-annotated images and weakly-annotated images as input.When the input is weakly-annotated images,an online refinement strategy is taken to find the proposal that cover the whole object rather than the most discriminative parts.A transfer training strategy for mix supervision is proposed in this thesis.The detector can get higher proposal recall for weakly-annotated data by making it has access to localization knowledge of the detector learned with fully-annotated data,which will lead to better detection performance.To further improve the performance,the weakly-labeled images are fed into the deep network in the order of from easy to complex;the process is formulated as curriculum learning.The proposed method is validated on the PASCAL VOC 2007 benchmark,and obtains 90% of a fully-supervised Faster R-CNN's performance(measured using mAP)with 15% of fully-supervised annotations together with image-level annotations for the rest images.The results show that the proposed learning framework can significantly reduce the labeling efforts for obtaining reliable object detectors.