Category-Aware Spatial Constraint For Weakly Supervised Object Detection

Visual object detection is sitting on the core of computer vision research.Recently with the development of deep convolutional neural networks,object detections based on deep learning have achieved great progresses.However,current the state-of-the-art object detection algorithms require training dataset with object-level label to learn the models.It is time-consuming and labor-intensive to obtain such object-level label and can introduce some annotator bias.This paper focus on problem of object detection based on weakly supervised learning,which is lack of object-level label and only have image-level in dataset for training the object detector.Weakly supervised object detection has wide application and important significance,and has seen a surge of attention from the computer vision community in the recent years.Current weakly supervised object detection algorithms are based on local,instance-level information.Therefore,we propose to exploit and incorporate unsupervised object global shape and location information to assist the model training.The follows are the main contents of this paper:1.We propose a novel category-specific pixel gradient map.During training,we extract category-specific pixel gradient map,based on which we roughly obtain location and shape of target object;2.We combine the rough estimation of object and location of regions to propose the spatial constraint,based on which we can introduce category-specific global information and region local information to the learning process;3.We propose a multi-center regularization to penalize violations between category centers and high-score regions,which makes the learning more stable.Our algorithm neither arises the complexity of model,nor requires addition supervised information to learning.Finally,extensive experiments show that our approach achieves outstanding performance of object detection and localization,and significantly outperforms the state-of-the-art methods.