DenseGAP: Graph-Structured Dense Correspondence Learning with Anchor Points
Preprint on Arxiv, 2021
Zhengfei Kuang1,2    Jiaman Li1,3    Mingming He2    Tong Wang2    Yajie Zhao2   
University of Southern California1    USC Institute for Creative Technologies2    Stanford University3   

Establishing dense correspondence between two images is a fundamental computer vision problem, which is typically tackled by matching local feature descriptors. However, without global awareness, such local features are often insufficient for disambiguating similar regions. And computing the pairwise feature correlation across images is both computation-expensive and memory-intensive. To make the local features aware of the global context and improve their matching accuracy, we introduce DenseGAP, a new solution for efficient Dense correspondence learning with a Graph-structured neural network conditioned on Anchor Points. Specifically, we first propose a graph structure that utilizes anchor points to provide sparse but reliable prior on inter- and intra-image context and propagates them to all image points via directed edges. We also design a graph-structured network to broadcast multi-level contexts via light-weighted message-passing layers and generate high-resolution feature maps at low memory cost. Finally, based on the predicted feature maps, we introduce a coarse-to-fine framework for accurate correspondence prediction using cycle consistency. Our feature descriptors capture both local and global information, thus enabling a continuous feature field for querying arbitrary points at high resolution. Through comprehensive ablative experiments and evaluations on large-scale indoor and outdoor datasets, we demonstrate that our method advances the state-of-the-art of correspondence learning on most benchmarks.


The following images show the concept and architecture design of our model.

Illustration of the concept. From left to right: 1) Showcase of how anchor points guide our model to find dense correspondences; 2) Visualization of our designed graph; 3) Three types of edges in our graph.

Overview of the framework. Given two images and anchor points, we first extract the coarse and fine feature maps of each image. Then we obtain the features of anchor points from the coarse feature maps as input to the Propagation Module. The output of the module is updated coarse feature maps, and is then fed with the fine feature maps to the Refinement Module.This module finally generates the updated fine feature maps.

Application 1: Outdoor Sparse Correspondence.

Sparse matching results on MegaDepth, compared with DualRC-Net and SuperGlue.

The reference image (1st column) is warped to the query image (2nd column) based on the dense correspondences generated by the baseline method (3rd) and our model (4th). The confidence maps of our predictions (represented by cycle consistency) are also shown in the5thcolumn. The black pixels in the confidence map represent those with the cycle consistency larger than 10 pixels.

Footer With Address And Phones