Abstract | ||
---|---|---|
Towards multi-class novelty detection, we propose an end-to-end trainable Parallel Network (PN) using no additional data but only the training set. Our key idea is to first divide the training set into successive subtasks of pseudo-novelty detection to simulate real scenarios. We then design PN which consists of multiple branches to address the pseudo-tasks. This yields a more compact and discrminative classification space and forms a natural ensemble. In practice, we sample a random subset of the training classes as pseudo-novel and regard the extra as known. The sampling result forms a sub-task fed to one branch of PN. All training classes are equally divided as pseudo-novel to PN's branches for better data balance and model diversity. By distinguishing between the known and the diverse pseudo-novel, PN extracts the concept of novelty in a compact classification space. This provides PN with generalization ability to real novel classes which are absent during training. During online inference, this ability is further strengthened with the ensemble of PN's multiple branches. Experiments on three public datasets show our method's superiority to the mainstream methods. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1109/ICPR48806.2021.9412234 | 2020 25TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR) |
DocType | ISSN | Citations |
Conference | 1051-4651 | 0 |
PageRank | References | Authors |
0.34 | 0 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Shuaiyuan Du | 1 | 0 | 1.69 |
Chaoyi Hong | 2 | 0 | 1.01 |
Zhiyu Pan | 3 | 0 | 1.35 |
Chen Feng | 4 | 61 | 6.14 |
Zhiguo Cao | 5 | 314 | 44.17 |