Core Concepts
KP-RPE enhances ViT models' robustness to unseen affine transformations by incorporating keypoint information.
Abstract
Geometric alignment is crucial for recognition tasks like face and gait recognition.
KP-RPE leverages key points to improve ViT models' resilience to scale, translation, and pose variations.
RPE introduces relative spatial relationships in ViTs, improving performance in unseen affine transformations.
KP-RPE dynamically adapts spatial relationships based on keypoints, enhancing model robustness.
Experimental results show significant improvements in face and gait recognition with KP-RPE.
Stats
RPE enables the model to capture relative spatial relationships among image regions.
Adding RPE increases performance in AffNIST test set.
KP-RPE adjusts spatial relationships based on keypoints for improved model adaptability.