Attention-based Dynamic Subspace Learners
Sukesh Adiga Vasudeva, Jose Dolz, Herve Lombaert
Deep metric learning methods are widely used to learn similarities in the data. Most methods use a single metric learner, which is inadequate to handle the variety of object attributes such as color, shape, or artifacts in the images. Multiple metric learners could focus on these object attributes. However, it requires a number of learners to be found empirically for each new dataset. This work presents a Dynamic Subspace Learners to dynamically exploit multiple learners by removing the need of knowing apriori the number of learners and aggregating new subspace learners during training. Furthermore, the interpretability of such subspace learning is enforced by integrating an attention module into our method, providing a visual explanation of the embedding features. Our method achieves competitive results with the performances of multiple learners baselines and significantly improves over the classification network in clustering and retrieval tasks.
Thursday 7th July
Poster Session 2.2 - onsite 11:00 - 12:00, virtual 15:20 - 16:20 (UTC+2)