Efficient Object Detection With Feature Sharing

Object detection, which aims at object localization and recognition, is a key com-ponent of computer vision. Although it has been studied for dozens of years, object de-tection is still immature facing the complicated real world. As a complex mix question of localization and recognition, object detection is always stuck with the tradeoff be-tween the computation cost and the model capacity. A central contribution of this thesis is exploring ways of reducing the intrinsic computation complexity of object detection and enhancing the model capacity. Focus on the topic of efficient object detection with feature sharing, we propose four novel algorithms, which are correlated to each other and applied on two detection tasks respectively. One is general object detection with powerful convolutional neural networks; the other is rapid face detection for mobile devices.We propose a novel object detection framework based on spatial pyramid pool-ing on convolutional neural network, which breaks the computational barrier between independent image regions, and enables the feature sharing between per-region classi-fiers in convolutional network based object detection system. This framework greatly improves the computational efficiency of per-region classification in object detection. Meanwhile, this framework makes any-size inputs for convolutional networks, provid-ing more flexibility. Extensive experiments show that this system accelerates the per-region classification in object detection by dozens of times.Based on above detection framework, we propose a novel object proposal network. Solving multi-scale/ratio detection with novel anchor pyramid, this network enables feature sharing in multi-scale detection. Further feature sharing between this object proposal network and the object detection network improves the efficiency of the entire detection system. The results on multiple general object detection benchmarks indicate that our system improve the performance and efficiency at the same time. As the first near real-time CNN based general object detection system, our framework has high impact on both research and real applications.Combining the two algorithms above, the proposed general object detection sys-tem significantly improves the detection accuracy, meanwhile achieves speedup of more than 200 times. Based on this detection system, we won the 1st place in the detection challenge of ILSVRC/ImageNet 2015, which is the largest and the most influential de-tection challenge in the world.We propose novel local binary feature and related learning method for face align-ment (precise face localization). The discriminative local feature learning and global feature shared regression, greatly enhance the model capacity and efficiency of the face alignment model. The proposed face alignment system is the faster alignment system nowadays. It achieves over 3,000 fps on a desktop or 300 fps on a mobile phone for locating a few dozens of landmarks.With the proposed binary feature, we come up with joint framework for face de-tection and alignment. In this framework, we learn shared feature between these two tasks, which greatly boosts the performance and reduces the overall computational and memory cost. The proposed face detection is one of the most efficient systems in the world.Working on the tasks of general object detection, and face detection, we propose framework for feature sharing between region classifiers and within each region clas-sifier, respectively. These kinds of feature sharing reduce the computation complexity, meanwhile enhance the model capacity and generalization. We believe the proposed efficient detection frameworks will boost related research and applications on object detection.