Experiments

Feature Extraction and Matching Algorithms

Learning-based feature extraction and matching algorithms like SuperPoint and SuperGlue outperform traditional methods such as SIFT and NN.

  • Both qualitative and quantitative experiments demonstrate the superior performance of learning-based methods, such as SuperPoint and SuperGlue (SP&SG), compared to traditional approaches like SIFT, adalma, and NN.
  • Besides the number of keypoints and matches, SP&SG is faster than SIFT&NN.

Intrinsics Refinement

We tested our intrinsics refinement on both COLMAP and Pixel-Perfect pipelines.

The following table shows intrinsics errors before refinement with ground-truth extrinsics:

The following table shows intrinsics errors after refinement with ground-truth extrinsics:

The following table shows intrinsics errors of PixelSfM before and after refinement with ground-truth extrinsics:

  • Before intrinsics refinement, PixelSfM yields much lower focal error and slightly higher principal point error than COLMAP. This is because featuremetric refinements improves the precision of keypoints.
  • After intrinsics refinement with GT extrinsics, our method yields much lower focal error and principal point error on both COLMAP and PixelSfM pipeline.
    • ~30% improvement on focal error.
    • >90% improvement on principal point error.

The following picture shows the L1 erro curve during optimization:

The principal point error continues decreasing during optimization, while the focal error slightly increases in the last several epoches. This is because the inherent noise contained in the keypoints. One method to mitigate it is to add featuremetric loss term in the object function, similar to Pixel-Perfect refinement, to reduce keypoint error.