University of Twente Student Theses
Novelty detection with Active Class-Incremental Learning on Long-tailed datasets
Lievense, M.M (2024) Novelty detection with Active Class-Incremental Learning on Long-tailed datasets.
|
PDF
4MB |
| Abstract: | In many practical applications, the effort required for data annotation often exceeds the cost of data acquisition, especially in the presence of long-tailed distributions within an open-set setting. This challenge is further exacerbated when rare classes must be identified in large, unlabeled datasets. Traditional methods operate under the closed-set assumption, which is frequently impractical in real-world scenarios. Existing research on open-set recognition does not fully capture the complexity of discovering novel classes during the annotation process. To address these limitations, we propose a methodology that integrates active learning and class-incremental techniques, utilizing out-of-distribution detection algorithms to efficiently identify novel classes during the annotation process. Our results demonstrate a significant reduction in the annotation effort required to approach a closed-set dataset on three widely used benchmark datasets. Specifically, our methodology discovers 100% of the classes on Places365-LT and ImageNet-LT with 59.1% and 57.6% fewer annotations, respectively, compared to random sampling, by employing a committee of multiple detectors. Similarly, we discover 99% of the classes on iNaturalist2018-Plantae with 23.0% fewer annotations. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Embedded Systems MSc (60331) |
| Link to this item: | https://purl.utwente.nl/essays/104449 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page