The competition results of the “Robust Vision Challenge” organized by the 2018 International Conference on Computer Vision and Pattern Recognition (CVPR) were unveiled on June 18 in Salt Lake City, USA. Under the joint guidance of Prof. Zhiguo Cao of HUST and Prof. Chunhua Shen of Adelaide University, doctoral student Ke Xian and master student Ruibo Li participated “Single Image Depth Prediction” competition and won the 2nd place in the world. On behalf of the team, Ke Xian received the certificate and gave a speech titled “Deep Attention-based Classification Network for Robust Depth Prediction”.
Different from the acquisition of scene distance information through expensive lidar, the depth prediction of single image aims at resuming the distance from each pixel to the imaging plane in a single image and provides the usual plane image signal with stereo information. Single image depth prediction is one of the basic technologies of robot stereo scene perception, unmanned system navigation, plane video to stereo video conversion, virtual reality panoramic information stereo, smartphone applications. It is also one of the frontier technologies of computer vision and has broad application foreground. In this competition, the single image depth prediction was different from conventional tasks. It asked for a set of model parameters that could be compatible with both indoor and outdoor modes, to test the algorithm’s adaptability to different scenarios. Then the competition is synthetically sorted according to various quantitative performance indexes finally achieved by the algorithm. This competition has attracted many well-known universities around the world, and it reflects the highest research level in the world in this field.
CVPR is the annual academic conference of IEEE and one of the three most recognized conferences in the field of computer vision in the world. It has strict criteria for paper acceptation., hence its citation and academic influence come top in computer science. This year, 3,309 valid contributions were submitted while only 979 papers were accepted eventually. The number of registered participants reached up to 6512, which shows the influence of the conference in the academic circle. It is worth mentioning that Ke Xian’s paper on image depth prediction “Monocular Relative Depth Perception with Web Stereo Data Supervision” was also accepted by the conference.
