Mutual Constraint Learning for Weakly Supervised Object Detection

The abundance of image-level labels and the lack of large scale bounding boxes detailed annotations promotes the expansion of Weakly-Supervised techniques for Object Detection (WSOD). In this work, we propose a novel mutual constraint learning for convolutional neural networks applied to detect bounding boxes only with global image-level supervision. The essence of our architecture is two new differentiable modules, Determination Network, and Parameterised Spatial Division, which explicitly allows the spatial division of the feature map within the network. These learnable modules give neural networks the ability to constructively generate shadow activation maps, dependent on the class activation maps. To demonstrate the effectiveness of our model for WSOD, we conduct extensive experiments on the multi-MNIST dataset. Experimental results show that mutual constraint learning can (i) help improve the accuracy of multi-category tasks, (ii) implement in an end-to-end way only with the image-level annotations, and (iii) output accurate bounding box labels, thereby achieving object detection.