Back to case studies

Robust Manhattan directions for panoramic images.

Panorama Vanishing Point Estimation

A spherical vanishing-point estimator for equirectangular panoramas, using cube-face line detection, weighted lookup tables, and Manhattan triplet fitting.

Vanishing PointsPanoramaSpherical LUTRANSAC

Detailed write-up →

Problem

Classic pinhole vanishing-point methods do not transfer directly to panoramas because image lines bend in equirectangular space and global spherical evidence is expensive to evaluate repeatedly.

Approach

  • Detect local perspective line segments after cube projection to avoid equirectangular line distortion.
  • Map line evidence onto a spherical parameter space and accumulate support in a lookup table.
  • Sample and score Manhattan direction triplets efficiently with constant-time candidate evaluation.

My Contribution

I independently designed and implemented the complete panoramic vanishing point estimation module, from algorithm to C++ implementation and evaluation.

On the algorithm side, I built the core modules:

  • Cube-face projection and line segment detection for distortion-free line extraction
  • Spherical weighted lookup table construction enabling constant-time RANSAC candidate evaluation
  • Manhattan triplet fitting under orthogonal constraints with brute-force sampling

Outcome

The module provides stable room orientation cues for downstream rectification, plane reasoning, and indoor reconstruction initialization.

Panorama vanishing point pipeline
Vanishing-point estimation pipeline.
Spherical lookup table evidence
Spherical evidence used for fast candidate scoring.