Multi-Resolution Generative Adversarial Networks for Tiny-Scale Pedestrian Detection

Although pedestrian detection techniques have achieved great success recently, the accurate tiny-scale pedestrian detection remains challenging with low resolution. The existing CNN based detection methods are limited in terms of in-line mechanism when encountering tiny-scale pedestrian detection, because the embedded convolution and pooling operations tend to weaken the features' representation of tiny scales objects. To ameliorate, we propose a newly-designed Multi-Resolution Generative Adversarial Network (MRGAN) to simultaneously conduct multi-resolution pedestrian detection by directly generating a high-resolution pedestrian image from low-resolution image. The key idea is to explore the intrinsic relations between high-resolution pedestrians and low-resolution pedestrians to enhance the representation of low-resolution pedestrians. The classification loss will be used to conduct the training process of super-resolution generative adversarial network, so that the generated super-resolution images can benefit the tiny-scale pedestrian detection. Besides, we also define a segmentation-based perceptual loss by incorporating a pre-trained image segmentation sub-network to refine the detail information. Extensive experiments and comprehensive evaluations on public challenging benchmarks confirm that our method outperforms the state-of-the-art methods.