Distilling Object Detectors via Decoupled Features

Knowledge distillation is a widely used paradigm for inheriting information from a complicated teacher network to a compact student network and maintaining the strong performance. Different from image classification, object detectors are much more sophisticated with multiple loss functions in which features that semantic information rely on are tangled. In this paper, we point out that the information of features derived from regions excluding objects are also essential for distilling the student detector, which is usually ignored in existing approaches. In addition, we elucidate that features from different regions should be assigned with different importance during distillation. To this end, we present a novel distillation algorithm via decoupled features (DeFeat) for learning a better student detector. Specifically, two levels of decoupled features will be processed for embedding useful information into the student, i.e., decoupled features from neck and decoupled proposals from classification head. Extensive experiments on various detectors with different backbones show that the proposed DeFeat is able to surpass the state-of-the-art distillation methods for object detection. For example, DeFeat improves ResNet50 based Faster R-CNN from 37.4% to 40.9% mAP, and improves ResNet50 based RetinaNet from 36.5% to 39.7% mAP on COCO benchmark. Code will be released^1,2.

1. Introduction

As one of the fundamental computer vision tasks, object detection has attracted increasing attention in various real-world applications including autonomous driving and surveillance video analysis. Recent advances of deep learning introduce many convolutional neural network based solutions to object detection. The backbone of a detector is often composed of heavy convolution operations to produce intensive features that is critical to the detection accuracy. But doing so inevitably results in a sharp increase in the cost of computing resource and an apparent decrease in detection speed. Techniques such as quantization [19], [58], [31], [57], [62], pruning [2], [17], [20], network design [55], [49], [15], [18] and knowledge distillation [56], [6] have been developed to overcome this dilemma and achieve an efficient inference on detection task. We are particularly interested in knowledge distillation [24], as it provides an elegant way to learn a compact student network when a performance proven teacher network is available. Classical knowledge distillation methods are firstly developed for the classification task to decide which category the image belongs to. The information from soft label outputs [24], [28], [38], [13] or intermediate features [1], [23], [66] of a well-optimized teacher network have been well exploited to learn the student networks, but these methods cannot be directly extended to the detection task which needs to further figure out where the objects are.