CVPR Daily - Tuesday

multiple tasks, including object detection, moving object detection, and semantic segmentation. “ One of the novel tasks we have is called lens soiling detection, ” Ganesh tells us. “ Fisheye cameras are usually placed on the external body of a camera, so they are susceptible to outside artefacts like mud and water droplets . Detecting disturbances on the lens is important because if this happens while driving, for example, you’ll want to turn off the automated driving feature and hand control back to the driver. However, for fisheye camera based autonomous driving, there is no publicly available dataset for research . WoodScape has over 10,000 images captured across Europe with different mud and water droplet patterns. The task seems easy, but in our experience, most of the self-supervised algorithms won’t help for various practical reasons. We’re excited to see what solutions the community come up with to solve this! ” To work out the baseline for the challenge, Ganesh used a network called PSPNet with a backbone of ResNet-50 . When he tried the network on the WoodScape Challenge dataset, his Intersect over Union (IoU) score was just 55 per cent, whereas the best participant has an IoU score of 83 per cent. “ That shows that standard CNNs are not good enough to work on wide field view camera data, ” Ganesh points out. “ Particularly the data that is captured under different geography and weather conditions, with all the diversity in that. Our top participants 16 DAILY CVPR Tuesday Workshop