Intrinsics Prediction Leaderboard

Task

In this task, participants are asked to submit camera intrinsics for the frames in our test set.

Metrics

Raw Intrinsics % Error: percent difference between predicted value and ground truth value for the following raw intrinsics parameters:

End-point-error (EPE): given a fixed set of 3D points, measures the average L2 distance between their 2D projections when using predicted versus ground truth camera intrinsics.

Submission Format

Please follow the instructions here to prepare and upload your intrinsics submission. Our benchmark currently supports two types of camera intrinsics models: rad-tan and mei. If you would like to submit predictions using an alternative camera model, please contact us at influxbenchmark@gmail.com.

Leaderboard

Use the "Sort by" dropdown to re-order the leaderboard based on the selected column.

Sort by:

Method	% fx Error ↓	% fy Error ↓	% cx Error ↓	% cy Error ↓	% Points Below 300 EPE ↑
COLMAP ^[1]	1274.785	1275.332	0.112	0.299	7.845
DroidCalib ^[2]	68.108	70.001	10.059	15.718	27.948
Perspective Fields ^[3]	64.618	64.608	18.572	19.660	17.766
GeoCalib ^[4]	56.537	56.527	0.099	0.204	52.883
AnyCalib_gen ^[5]	44.245	44.148	0.377	0.584	70.628
UniDepthV2 ^[6]	50.567	51.073	1.610	2.582	46.087
WildCamera ^[7]	45.561	46.901	5.040	6.390	47.192

[1] Structure-from-Motion Revisited. [paper] [code]

[2] Deep Geometry-Aware Camera Self-Calibration from Video. [paper] [code]

[3] Perspective Fields for Single Image Camera Calibration. [paper] [code]

[4] GeoCalib: Learning Single-image Calibration with Geometric Optimization. [paper] [code]

[5] AnyCalib: On-Manifold Learning for Model-Agnostic Single-View Camera Calibration. [paper] [code]

[6] UniDepthV2: Universal Monocular Metric Depth Estimation Made Simpler. [paper] [code]

[7] Tame a Wild Camera: In-the-Wild Monocular Camera Calibration. [paper] [code]