Abstract: Occlusion is a common phenomenon in images characteristic of complex scenes. Discovering the pattern of how occlusion affects image cognition and establishing a cognition model insusceptible to occlusion is closely related to the utilization of computer vision technologies, and it is also an important and pressing scientific problem to be solved. By analyzing occlusion in complex scenes and how occlusion affects image cognition, this paper established the Tsinghua University Occlusion Image Database for evaluating the anti-occlusion performance of algorithms and studying image cognition model. Firstly, based on occlusion’s impact on image cognition, this paper proposed a 4-dimension occlusion attribute including occluded part, occluded area, occlusion relationship and occlusion complexity, as well as a quantification standard for the extent of occlusion. Then we proposed a novel hierarchical dataset structure, based on which the database could be constructed. This paper established TUOD database which consists of 2 100 images extracted from PASCAL VOC and Image Net databases. Those images covered 4 major object types: aeroplane, car, person and animal. An experiment was conducted to analyze the influence of each dimension of occlusion attribute on the performance of Faster R-CNN using images in TUOD. As is shown in the aforementioned experiment, TUOD database can provide quantitative criteria for the anti-occlusion performance of algorithms and thus it is highly practical and lays the foundation for improving the anti-occlusion performance of object recognition algorithms in complex scenes.

Key words:  influence of occlusion classification, assessment of anti occlusion capability, assessment of anti occlusion capability, occlusion image database