ZERO – Detect objects without training examples by knowing their parts
conference paper
Current object recognition techniques are based on deep learning and require substantial training samples in order to achieve a good performance. Nonetheless, there are many applications in which no (or only a few) training images of the targets are available, whilst they are well-known by domain experts. Zero-shot learning is used in use cases with no training examples. However, current zero-shot learning techniques mostly tackle cases based on simple attributes and offer no solutions for rare, compositional objects such as a new product, or new home-made weapons. In this paper we propose ZERO: a zero-shot learning method which learns to recognize objects by their parts. Knowledge about the object composition is combined with state-of-the-art few-shot detection models, which detects the parts. ZERO is tested on the example use case of bicycle recognition, for which it outperforms few-shot object detection techniques. The object recognition is extended to detection by localizing it, by taking into account knowledge about the object’s composition, of which the results are studied qualitatively
TNO Identifier
955835
Source title
Proceedings of the AAAI 2021 Spring Symposium on Combining Machine Learning and Knowledge Engineering, AAAI-MAKE 2021, Palo Alto, CA, 22-24 March 2021
Files
To receive the publication files, please send an e-mail request to TNO Repository.