Relative Depth Estimation from Dense Feature Matches

Vladimir Salnikov · BSc thesis, Lund University · CVML group

My Bachelor’s thesis, written in the CVML group at Lund University under Viktor Larsson.

Problem

Given two images of the same scene and a set of dense feature matches between them, can we estimate relative depth at the matched points — i.e. for each match, which of the two corresponding 3D points is closer to the camera — without running full multi-view stereo?

Approach

The thesis trains two complementary families of models on this signal:

• Graph Neural Networks over the match graph, with each match as a node and edges encoding geometric / appearance similarity between matches.
• Convolutional Neural Networks working directly on local image patches around each match.

Both are compared against existing relative-depth methods from the CVML literature. Implementation in PyTorch + PyTorch Geometric, with feature extraction and matching via OpenCV.

Reflections

The project was my first sustained piece of computer-vision research and the reason I’m continuing in visual computing at DTU. A lot of the work was infrastructure — preparing match datasets, sanity-checking baselines, comparing across published methods with PhD students in the group — but the comparisons themselves taught me more than any paper-reading would have.