YADA: you always dream again for better object detection

Abstract

Object detection has been attracting a lot of attention from the computer vision community. It has a wide range of practical applications ranging from the traditional use such as image annotation to modern uses such as self-driving vehicles, robotics, surveillance systems, and augmented reality. Recently, deep learning has significantly improved the state-of-the-art performance of the object detection task. Many works explore various deep network structures to improve the performance. However, the impact of training data is still not well investigated. Although some works focus on data augmentation and data synthesis, there is no guarantee that they are effective for the training process. In this paper, we propose a novel framework addressing the problem of generating relevant data and how to use them effectively. We apply lucid data synthesizing which generates data by mining hard examples and embedding them to the same context locations. Further, we utilize a dual-level deep network leveraged with these generated data to effectively detect hard objects in images. Extensive experiments on two benchmarks, PASCAL VOC and KITTI, demonstrate the superiority of our approach over the state-of-the-art methods.