From Robustness to Efficiency: Deformation-Aware and Efficient Local Feature Extraction for Images

Just as animals rely on visual perception and geometric understanding to navigate the 3D world, modern computers emulate this ability through Simultaneous Localization and Mapping (SLAM), image-based 3D reconstruction, and visual place recognition techniques, all relying on image features for obtaining correspondences. However, most feature extraction methods handle only affine transformations, ignoring non-rigid deformations, ubiquitous in the real-world. This work investigates deformation-aware local features, leveraging RGB-D images to compute geodesics, where RGB denotes the visible channels (Red, Green, Blue) and D represents image depth. Then, we generalize the concept to RGB-only images via learned representations. We introduce a novel RGB-D dataset with non-rigid deformations for real-world benchmarking, where experiments showed significant improvements in foundational vision tasks as matching and registration when adopting our proposed strategies. Finally, we present an efficient local feature extractor, balancing accuracy with reduced computational cost, expanding visual perception for mobile computers.