Integrating Dashcam Views through Inter-Video Mapping
Wei-Tse Lee
Fan Wang
National Taiwan University
In IEEE International Conference on Computer Vision (Proceedings of ICCV 2015)
An overview of our inter-video mapping approach. Given $I^R$ and $I^T$ as input (a), for each incoming image of $I^T$, we seek its visually closest image in $I^R$ as the bridge image. To estimate the dense mapping function for perspective adaptation between the reference and the target image, we perform long-range motion estimation between the reference and the bridge image, together with the estimation of a non-linear geometric transformations that related the image-coordinates of the bridge and the target image (b). The projected shape and size of the reference image is adjusted and stitched into the occluded region to generate the final composited image (c).
In this paper, an inter-video mapping approach is proposed to integrate video footages from two dashcams installed on a preceding and its following vehicle to provide the illusion that the driver of the following vehicle can see-through the preceding one. The key challenge is to adapt the perspectives of the two videos based on a small number of common features because most of the common region of the video captured by the following vehicle is occluded by the preceding one. The proposed inter-video mapping, which estimates the spatially-varying motions in the space-time path within the video captured by the preceding vehicle and between the two videos. Specifically, we gradually collect the correspondences from local trajectories and incrementally propagate them to form long-range motion estimation. This new formulation enables accurate alignment of frames with large geometric variations in a video. To dynamically infer spatially-variant warps across the two videos, locally-grouped features are utilized to perform trajectory transfer to maintain temporal coherence. The two procedures are tightly coupled, resulting in the observed perspective discrepancy between the two videos. Besides, once the inter-video mapping is established, the correspondences can be updated incrementally, so the proposed method is suitable for on-line applications.
technical paper
presentation slide

     author = {Hsin-I Chen and Yi-Ling Chen and Wei-Tse Lee and Fan Wang Bing-Yu Chen},
     title = {Integrating Dashcam Views through Inter-Video Mapping},
     booktitle = {IEEE International Conference on Computer Vision},      year = {2015},
     pages = {},
1504 visitors since August 2013