Centroid- and Orientation-aware Feature Learning
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: disentanglement, disentangling, rotational equivariant convolutional layer, centroid learning, orientation learning, feature learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: The proposed method provides a better way to learn centroids, orientations and their invariant features of objects
Abstract: Robust techniques for learning centroids and orientations of objects and shapes in two-dimensional images, along with other features is crucial for image- and video-processing applications. While this has been partially addressed using a number of techniques by achieving translational and rotational equivariance and invariance properties, learning them as part of the features still remains an open problem. In this paper, we propose a novel encoder-decoder-based mechanism for learning independent factors of variations, including centroids and orientations, by embedding special layers to achieve translational and rotational equivariance and invariance. Our evaluation, across a number of datasets, including that of real-world ones, against five different state-of-the-art baseline models shows that our model not only can offer superior disentangling and reconstruction performance, but also offers exceptional training and inference performance, as much as 10X for training and 9X on inference compared to the average performance of other models.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2862
Loading