ViewSynth: Learning Local Features from Depth using View Synthesis

TL;DR

ViewSynth introduces a novel framework for learning depth image features by jointly training view synthesis and keypoint descriptors, significantly improving 3D matching and localization performance.

Contribution

It proposes a new method that explicitly reasons beyond visible pixels by synthesizing views, enhancing depth feature learning for keypoint detection and matching.

Findings

Outperforms state-of-the-art in 3D keypoint matching and camera localization.

Demonstrates strong generalization across different RGB-D datasets.

Effectively encodes occluded scene information through view synthesis.

Abstract

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.