In this vertical we investigate different self-supervised and unsupervised learning
schemes for scene and object understanding.
a. Unsupervised Representation Learning
In this project, a challenging problem of unsupervised discovery of object
landmarks is addressed with a focus on improving the limitations of existing
methods. We propose a consistency-guided bottleneck in an image reconstruction-
based pipeline, leveraging landmark consistency and pseudo-supervision to
generate adaptive heatmaps. Landmark correspondence across images is utilized
for pseudo-supervision, and land-mark consistency is employed to modulate the
uncertainty of discovered landmarks.
Related Publications
● Unsupervised Landmark Discovery Using Consistency-Guided Bottleneck, BMVC, 2023
● Pose-Guided Self-Training with Two-Stage Clustering for Unsupervised Landmark Discovery, CVPR, 2024
b Learning Structure-Aware Deep Spectral Embedding
In this project, we focus on the learning of structure-aware deep spectral
embedding. The current challenge lies in the preservation of the subspace
structure of data from the original space to the embedding space.
Related Publications
c. Constrained Metric Learning by Permutation Inducing Isometries
In this study, the focus is on metric learning algorithms that attempt to improve
performance by learning a more appropriate metric. The exploration involves
constraining the learned metric to be invariant to the geometry-preserving
transformations of images that induce permutations in the feature space.
Related Publications
d. Data Augmentation
Related Publications
● DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models, CVPR, 2024.