Matching with AffNet based rectifications

TL;DR

This paper introduces two novel methods, DenseAffNet and DepthAffNet, for two-view matching under significant viewpoint changes, improving speed and accuracy by leveraging affine shape estimates and depth information.

Contribution

The paper presents two new methods that reduce view synthesis overhead and enhance matching accuracy under large viewpoint variations.

Findings

DenseAffNet is faster and more accurate on generic scenes.

DepthAffNet performs comparably to state-of-the-art on scenes with large planes.

Both methods are validated on three public datasets.

Abstract

We consider the problem of two-view matching under significant viewpoint changes with view synthesis. We propose two novel methods, minimizing the view synthesis overhead. The first one, named DenseAffNet, uses dense affine shapes estimates from AffNet, which allows it to partition the image, rectifying each partition with just a single affine map. The second one, named DepthAffNet, combines information from depth maps and affine shapes estimates to produce different sets of rectifying affine maps for different image partitions. DenseAffNet is faster than the state-of-the-art and more accurate on generic scenes. DepthAffNet is on par with the state of the art on scenes containing large planes. The evaluation is performed on 3 public datasets - EVD Dataset, Strong ViewPoint Changes Dataset and IMC Phototourism Dataset.