Risky Region Localization with Point Supervision

Kazuki Kozuka, Juan Carlos Niebles

Abstract: We study the problem of localizing regions in an image that depict potentially risky areas. In particular, we focus on images acquired by a front camera mounted on a car with the goal of localizing image regions where pedestrians are likely to enter the scene suddenly. In this case, we define the risk value at every pixel as the likelihood that a pedestrian will occupy those pixels shortly. This task is very challenging because the risk areas are not easily characterized by appearances of single objects, and therefore these regions exhibit large visual variations. Additionally, the boundaries of the risk regions in the image are not easily defined by human annotators, as they do not tend to correspond to object boundaries. This causes the annotation process to be ambiguous and costly. To overcome the ambiguity in the boundaries of risky regions, we adopt a weakly supervised method for risk region localization and risk value estimation that only requires single point supervision at training time. To evaluate our approach, we augment the Caltech Pedestrian dataset with risk region annotations. Our results show that our weak supervised method outperform fully supervised approaches in risk region localization and risk value estimation.