Hierarchical Scene Coordinate Classification and Regression for Visual Localization

Abstract

Visual localization is critical to many applications in computer vision and robotics. To address single-image RGB localization, state-of-the-art feature-based methods match local descriptors between a query image and a pre-built 3D model. Recently, deep neural networks have been exploited to regress the mapping between raw pixels and 3D coordinates in the scene, and thus the matching is implicitly performed by the forward pass through the network. However, in a large and ambiguous environment, learning such a regression task directly can be difficult for a single network. In this work, we present a new hierarchical scene coordinate network to predict pixel scene coordinates in a coarseto-fine manner from a single RGB image. The network consists of a series of output layers with each of them conditioned on the previous ones. The final output layer predicts the 3D coordinates and the others produce progressively finer discrete location labels. The proposed method outperforms the baseline regression-only network and allows us to train single compact models which scale robustly to large environments. It sets a new state-of-the-art for singleimage RGB localization performance on the 7-Scenes, 12-Scenes, Cambridge Landmarks datasets, and three combined scenes. Moreover, for large-scale outdoor localization on the Aachen Day-Night dataset, our approach is much more accurate than existing scene coordinate regression approaches, and reduces significantly the performance gap w.r.t. explicit feature matching approaches.

Visual Camera Re-Localization from RGB and RGB-D Images Using DSAC

“…Li et al [52] improved on this initial effort by enforcing multi-view and photometric consistency throughout training. In a follow-up work, Li et al [54] introduce a joint classification-regression network architecture for predicting scene coordinate, and demonstrate the effectiveness of training data augmentation for large improvements on standard benchmarks.…”

“…We convert input images to grayscale and re-scale them to 480px height. For training, we follow Li et al [54] and apply data augmentation. We apply random adjustments of brightness and contrast of the input image within a ±10% range.…”

“…We randomly re-scale images within 66% and 150%, and adjust the focal length accordingly. Different from Li et al [54], we do not shear training images, since our simple pinhole camera model does not support this operation. We also do not shift images.…”

RnR: Retrieval and Reprojection Learning Model for Camera Localization

“…In 2019, they proposed an improved version named ngdsac [18], which optimizes arbitrary tasks loss during training process by allowing the network to learn the probability of each scene coordinate at the same time, thereby improving accuracy. Li et al proposed to use a hierarchical scene coordinate network [48] and predict pixel scene coordinates in a coarseto-fine manner, and significantly reduce the performance gap with feature matching methods. In general, although these methods achieve high accuracy in small scenes, they are still not scalable enough for large scenes due to the limitations of space projection methods.…”

Lost in the Woods? Place Recognition for Navigation in Difficult Forest Environments

“…In the case of Kendall et al ( 2015 ), Walch et al ( 2017 ), and more recently Li et al ( 2019 ), the output can be as complex as a 6 degree of freedom pose regression respective to the original “map” of images. What all of these systems have in common, however, is that the domain against which they can match novel images is fixed at the point at which they are trained.…”