Study on robust object detection considering multiple visual appearance of object

Abstract

In the real-world environment, we encounter objects of various appearances. Even in such environments, it is essential to design networks that are robust to the environments. In a single modality such as color or thermal, the visual appearances of an object are different due to occlusion, small-scale, etc. For robust object detection in the single modality, we analyze the concept of uncertainty to guide the detection network can focus on occlusion and small-scale objects. In addition, by focusing on the small-scale issue, a method for robust detection in the small-scale appearance was also introduced. We proposed a memory structure that mimics the cued recall process in which humans associate large-scale objects even with small-scale objects. Recently, studies extending to multispectral modality are being attempted. In such a multispectral modality, the uncertainty is introduced to effectively capture the appearance information of objects observed in multiple spectrums, and we design a multispectral object detector by considering the importance of each modality. In addition, in actual various applications, various visual appearances of objects can be used in the training phase, but it is not possibiel in the inference phase. We introduce various object detectors that can detect objects even when such information cannot be used in the inference phase. By doing so, we present a new perspective in the field of computer vision dealing with various aspects of objects.