Detecting small objects is notoriously challenging due to their lowresolution and noisy representation. Existing object detection pipelinesusually detect small objects through learning representations of all theobjects at multiple scales. However, the performance gain of such ad hocarchitectures is usually limited to pay off the computational cost. In thiswork, we address the small object detection problem by developing a singlearchitecture that internally lifts representations of small objects to"super-resolved" ones, achieving similar characteristics as large objects andthus more discriminative for detection. For this purpose, we propose a newPerceptual Generative Adversarial Network (Perceptual GAN) model that improvessmall object detection through narrowing representation difference of smallobjects from the large ones. Specifically, its generator learns to transferperceived poor representations of the small objects to super-resolved ones thatare similar enough to real large objects to fool a competing discriminator.Meanwhile its discriminator competes with the generator to identify thegenerated representation and imposes an additional perceptual requirement -generated representations of small objects must be beneficial for detectionpurpose - on the generator. Extensive evaluations on the challengingTsinghua-Tencent 100K and the Caltech benchmark well demonstrate thesuperiority of Perceptual GAN in detecting small objects, including trafficsigns and pedestrians, over well-established state-of-the-arts.
展开▼
