University of Twente Student Theses
Knowing the Unknown : Open-World Recognition for Biodiversity Datasets
Gangireddy, Rajesh (2023) Knowing the Unknown : Open-World Recognition for Biodiversity Datasets.
|
PDF
17MB |
| Abstract: | The world is a vast and mysterious place, teeming with countless unknowns. Computer vision models when deployed in real-world applications often encounter images that belong to categories that the models have not been trained to recognise. For a computer model to be reliable and open-world ready, it must be ‘aware’ of the open world and be able to recognise and reject an image from an ‘unknown’ category. In this research work, several existing research gaps in the domain of open-world recognition or out-of-distribution detection (OOD) are identified and addressed. A set of eight retrofittable OOD detection methods are evaluated on long-tailed and fine-grained biodiversity datasets. Furthermore, the concept of using domain similarity scores between datasets to quantify the difficulty in OOD detection is introduced. Additionally, limitations of existing state-of-the-art OOD detection methods were identified and the Entropy weighted nearest neighbour’s distance (EnWeDi) method that overcomes these limitations is proposed. The proposed method outperforms the existing methods and achieves the highest OOD accuracies in almost every experimental setup. The effect of image feature embeddings stacked from intermediate layers (of a CNN) is investigated in this work and ‘AutoCrop’ - a robust way of stacking feature embeddings for the task of OOD detection is proposed. Overall this research provides valuable insights into OOD detection and contributes to taking a step closer to the ultimate goal of making computer vision models more ‘openworld ready’ and for AI systems that use them to be reliable, trustable and safe. |
| Item Type: | Essay (Master) |
| Clients: | Intel Benelux BV, Groningen, Netherlands |
| 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/94784 |
| 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