Classical object recognition has made great progress in recent years, where the number of categories and training samples are fixed. To simulate human’s continual learning mechanism, the task of incremental object recognition has drawn increasing attention from both academia and industry.
CLVision workshop of CVPR2020(https://sites.google.com/view/clvision2020/overview) launched the first continual learning challenge in CVPR history. The organizers formalize three incremental problems from real-world scenarios and set up three corresponding challenge tracks: New Instances (NI), Multi-Task New Classes (MT-NC), New Instances and Classes (NIC). As for evaluation, apart from the most traditional classification accuracy, more metrics such as memory usage, disk usage, training time etc. are adopted. The pre-selection was performed on CodaLab and 79 teams registered for the challenge. Among them, the top 11 were chosen for final evaluation including teams from University of Toronto, Tokyo Institute of Technology, University of Bristol, University of Waterloo, New York University, Amazon, etc.
Team ICT_VIPL from our lab (student members: Chen He, Qiyang Wan and Fengyuan Yang) participated all the three tracks and was awarded Winner of NI track. Moreover, the team achieved the highest score in terms of the metric “classification accuracy” under the setting of overall three tracks. Chen He on behalf of the team has attended the workshop online and given a 5-min live presentation about our algorithm.
Fig. 1. Winner’s certificate on NI track
Fig. 2. Live presentation by Chen He