ViewSynth: Learning Local Features from Depth using View Synthesis
This addresses the challenge of adapting RGB local feature learning to depth images for computer vision applications, offering improvements in tasks like camera localization, though it is incremental by building on existing depth and RGB techniques.
The paper tackled the problem of keypoint detection and matching in depth images by proposing ViewSynth, a framework that jointly learns viewpoint-invariant features and view synthesis, outperforming state-of-the-art methods in 3D keypoint matching and camera localization on datasets like 7-Scenes, TUM RGBD, and CoRBS.
The rapid development of inexpensive commodity depth sensors has made keypoint detection and matching in the depth image modality an important problem in computer vision. Despite great improvements in recent RGB local feature learning methods, adapting them directly in the depth modality leads to unsatisfactory performance. Most of these methods do not explicitly reason beyond the visible pixels in the images. To address the limitations of these methods, we propose a framework ViewSynth, to jointly learn: (1) viewpoint invariant keypoint-descriptor from depth images using a proposed Contrastive Matching Loss, and (2) view synthesis of depth images from different viewpoints using the proposed View Synthesis Module and View Synthesis Loss. By learning view synthesis, we explicitly encourage the feature extractor to encode information about not only the visible, but also the occluded parts of the scene. We demonstrate that in the depth modality, ViewSynth outperforms the state-of-the-art depth and RGB local feature extraction techniques in the 3D keypoint matching and camera localization tasks on the RGB-D datasets 7-Scenes, TUM RGBD and CoRBS in most scenarios. We also show the generalizability of ViewSynth in 3D keypoint matching across different datasets.