Data-Driven Street Scene Layout Estimation for Distant Object Detection

We present a street scene layout estimation method based on transferring layout annotation from a (large) image database and its application for distant object detection. Inspired by nonparametric scene labeling approaches, we estimate a scene's geometric layout by matching global image descriptors and retrieving the most similar layout configuration. Our label transfer is done for each sub-region of an image and a tiered scene model is used to integrate all the local label information into a coherent scene layout prediction. Given the geometric layout, we use a super-resolution method to zoom in the distance region and refine the search in object detection. On KITTI dataset, we show that we can reliably generate scene layout and improve the detection of distant cars over the state of the art DPM detector.