DSAC — Differentiable RANSAC for Camera Localization

Abstract

RANSAC is an important algorithm in robust optimization and a central building block for many computer vision applications. In recent years, traditionally hand-crafted pipelines have been replaced by deep learning pipelines, which can be trained in an end-to-end fashion. However, RANSAC has so far not been used as part of such deep learning pipelines, because its hypothesis selection procedure is non-differentiable. In this work, we present two different ways to overcome this limitation. The most promising approach is inspired by reinforcement learning, namely to replace the deterministic hypothesis selection by a probabilistic selection for which we can derive the expected loss w.r.t. to all learnable parameters. We call this approach DSAC, the differentiable counterpart of RANSAC. We apply DSAC to the problem of camera localization, where deep learning has so far failed to improve on traditional approaches. We demonstrate that by directly minimizing the expected loss of the output camera poses, robustly estimated by RANSAC, we achieve an increase in accuracy. In the future, any deep learning pipeline can use DSAC as a robust optimization component 1 .

Full-Frame Scene Coordinate Regression for Image-Based Localization

“…Our CNN based approach presented in this paper is inspired by the state-of-the-art RGB-only DSAC pipeline [3], and it consists of two stages. In the first stage, a Coordinate CNN is adopted to generate 2D-3D correspondences from a given RGB-only image.…”

“…The Coordinate CNN used in the original DSAC pipeline [3] is trained and evaluated in a patch-based manner. Similar to the patch-based approaches for semantic segmentation, patches need to be sampled during both training and test time.…”

“…In the most recent DSAC paper [3], a neural network based SCoRF pipeline for RGB-only images is proposed. In contrast to previous SCoRF pipelines, two CNNs are adopted for predicting scene coordinates and for scoring hypotheses respectively.…”

“…However, their localization accuracy is still far below traditional approaches in other situations where local features perform well. Unlike PoseNet, approaches such as DSAC [3] obtain the 6 DoF pose by a two-stage pipeline: first, regressing a 3D point position for each pixel in an image, and then determining the camera pose using RANSAC based on these correspondences as in the conventional localization pipeline. Results have shown that these methods are the current state-of-the-art on the standard 7-Scenes benchmark [28].…”

“…This fullframe Coordinate CNN is trained to generate dense scene coordinate predictions from a given RGB image. Compared to the patch-based Coordinate CNN adopted in the DSAC pipeline [3], which is trained to generate a single scene coordinate prediction from an image patch centered around a pixel, our network considers more global information and can be evaluated efficiently at test time. Nevertheless, encoding more global context during regression is more prone to overfitting than using local patches since local patches are more invariant to viewpoint changes.…”