Unsupervised Object Detection using Patch Based Image Classifier and Gradient Importance Map

Image classification in computer vision has seen tremendous amount of success in recent years. Deep learning has played a pivotal role in achieving human level performance in many image recognition challenges and benchmarks. Even though, image classification has been so successful, no other closely related domains have taken advantage from the efforts put into development of image classification methods. One such closely related field is of Object Detection. Object detection or localisation is a computer vision problem whose solutions have not been victorious enough to human level performance. Many challenges arise when developing object detection models for newly generated domains, one of which is labelling of datasets. Preparation of dataset is one of the most cumbersome and expensive task to accomplish while developing an object detection model. Although, image classifiers are used as a feature extractor in object detection training regimes, their localisation abilities are barely studied. In this paper, we propose an object detection training regime, that does not rely on bounding box labelled datasets, hence unsupervised in nature, and is solely based on trained image classifiers. We build up on our hypothesis, that, "if an image classifier is able to predict what object is in the input image, then it must have information about where the object is, we just need a mechanism to extract that information from it". Precisely, we divide the input image into patches of same size and employ a parameter restricted convolutional classifier on each patches to predict whether it contains the object or not, we call this our patch-based image classifier (the object here is the prediction of the trained image classifier). The training of the patch-based classifier is not straightforward as there is no true labels for each patches on which we can reduce the binary cross-entropy. Therefore, we propose a loss function weighted by the importance map, which we generate using Grad-CAM, that when minimized detects patches only containing objects that the classifier predicted. Empirically, we show that our method performs competitively on the ILSVRC-2017 object localisation benchmark.