Advancements in Unsupervised Feature Learning for Image Data
Uttam Karki, Subarna Shakya · 2023
This research addresses the challenge of unsupervised landmark learning in the presence of non-parameter data, where image pairs exhibit significant variations in appearance and posture. Traditionally, unsupervised landmark learning relies on synthetic image pairs with similar appearances but differing postures. However, these synthetic pairs, generated through predefined transformations, do not accurately represent real-world appearance and pose changes. To overcome this limitation, we propose a novel approach incorporating Cross-Picture Cycle Compatibility (C3). We iteratively apply the technique of reconstructed exchange twice to establish comprehensive supervision, with a primary focus on maintaining cycle consistency within the framework. Additionally, we introduce a cross-picture flow to ensure the consistency of image landmarks.