# A tour of mrcal: triangulation

$$\DeclareMathOperator*{\argmin}{argmin} \DeclareMathOperator*{\Var}{Var}$$

This is an overview of a more detailed discussion about triangulation methods and uncertainty.

## Previous

We just looked at dense stereo processing

## Overview

We just looked at dense stereo processing, where we generated a range map: an image where each pixel encodes a range along each corresponding observation vector. This is computed relatively efficiently, but computing a range value for every pixel in the rectified image is still slow, and for many applications it isn't necessary. The mrcal triangulation routines take the opposite approach: given a discrete set of observed features, compute the position of just those features. In addition to being far faster, the triangulation routines propagate uncertainties and provide lots and lots of diagnostics to help debug incorrectly-reported ranges.

mrcal's sparse triangulation capabilities are provided by the mrcal-triangulate tool and the mrcal.triangulate() Python routine. Each of these ingests

• Some number of pairs of pixel observations $$\left\{ \vec q_{0_i}, \vec q_{1_i} \right\}$$
• The corresponding camera models
• The corresponding images

To produce

• The position of the point in space $$\vec p_i$$ that produced these observations for each pair
• A covariance matrix, reporting the uncertainty of each reported point $$\Var \vec p_i$$ and the covariances $$\Var \left( \vec p_i, \vec p_j \right)$$

## Triangulation

Let's use our Downtown Los Angeles images. Before we start, one important caveat: there's only one camera, which was calibrated monocularly. The one camera was moved to capture the two images used to triangulate. The extrinsics were computed with a not-yet-in-mrcal tool, and mrcal cannot yet propagate the calibration noise in this scenario. Thus in this example we only propagate the observation-time noise.

Image from the left camera:

Let's compute the range to the top of the "Paul Hastings" tower, near the center of the image. I'm looking at the "Paul Hastings" logo, roughly 566m (according to the map) from the camera. I have a pixel coordinate on the logo: $$\vec q = (2874, 1231)$$, the two images and the two models. This is enough information to triangulate:

\$ mrcal-triangulate                       \
--range-estimate 566                  \
--q-observation-stdev 0.3             \
--q-observation-stdev-correlation 0.1 \
--stabilize-coords                    \
--template-size 31 17                 \
--viz uncertainty                     \
splined-.cameramodel              \
.jpg                              \
2874 1231


Here we used the splined models computed earlier. We gave the tool the true range (566m) to use as a reference. And we gave it the expected observation noise level: 0.3 pixels (a loose estimate that should be roughly correct). We declared the left-camera/right-camera pixel observations to be correlated with a factor of 0.1 on the stdev, so the relevant cross terms of the covariance matrix are $$(0.3*0.1 \mathrm{pixels})^2$$. It's not yet clear how to get the true value of this correlation, but we can use this tool to gauge its effects.

The mrcal-triangulate tool finds the corresponding feature in the second image using a template-match technique in mrcal.match_feature(). This operation could fail, so a diagnostic visualization can be requested by passing --viz match. This pops up an interactive window with the matched template overlaid in its best-fitting position so that a human can validate the match. The match was found correctly here. We could also pass --viz uncertainty, which shows the uncertainty ellipse. Unless we're looking very close in, this ellipse is almost always extremely long and extremely skinny. Here we have: So looking at the ellipse usually isn't very useful, and the value printed in the statistics presents the same information in a more useful way. The mrcal-triangulate tool produces lots of reported statistics:

## Feature [2874. 1231.] in the left image corresponds to [2832.473 1234.974] at 566.0m
## Feature match found at [2831.603 1233.882]
## q1 - q1_perfect_at_range = [-0.87  -1.092]
## Triangulated point at [ -41.438 -166.48   476.58 ]; direction: [-0.082 -0.328  0.941]
## Range: 506.52 m (error: -59.48 m)
## q0 - q0_triangulation = [-0.022  0.548]
## Uncertainty propagation: observation-time noise suggests worst confidence of sigma=23.94m along [ 0.084  0.329 -0.941]
## Observed-pixel sensitivity: 56.72m/pixel (q1). Worst direction: [0.99  0.139]. Linearized correction: 1.05 pixels
## Calibration yaw (rotation in epipolar plane) sensitivity: -2095.34m/deg. Linearized correction: -0.028 degrees of yaw
## Calibration yaw (cam0 y axis)                sensitivity: -1975.17m/deg. Linearized correction: -0.030 degrees of yaw
## Calibration pitch (tilt of epipolar plane) sensitivity: 260.09m/deg.
## Calibration translation sensitivity: 238.47m/m. Worst direction: [0.986 0.    0.165]. Linearized correction: 0.25 meters of translation
## Optimized yaw   (rotation in epipolar plane) correction = -0.023 degrees
## Optimized pitch (tilt of epipolar plane)     correction = 0.030 degrees
## Optimized relative yaw (1 <- 0): -1.364 degrees


We see that

• The range we compute here is 506.52m, not 566m as desired
• There's a vertical shift 0.548 pixels between the triangulated point and the observation in the left camera: the epipolar lines aren't quite aligned, which means the calibration is a bit off. Either in the intrinsics or the extrinsics
• With the given observation noise, the 1-sigma uncertainty in the range is 23.94m, almost exactly in the observation direction. This is smaller than the actual error of 59.48m, which could be explained by any of
• the extrinsics were computed using the intrinsics, without taking into account the noise in the intrinsics; the extrinsics were then assumed perfect, since we're not propagating calibration-time noise
• the intrinsics are a bit off: we saw patterns when computing the intrinsics in the earlier, which would cause a bias
• the values of --q-observation-stdev and --q-observation-stdev-correlation weren't selected in a principled way, and could be off
• Moving the matched feature coordinate in the right image affects the range at worst at a rate of 56.72 m/pixel. Unsurprisingly, the most sensitive direction of motion is left/right. At this rate, it would take 1.05 pixels of motion to "fix" our range measurement
• Similarly, we compute and report the range sensitivity of extrinsic yaw (defined as the rotation in the epipolar plane or around the y axis of the left camera). In either case, an extrinsics yaw shift of 0.03 degrees would "fix" the range measurement.
• We also compute sensitivities for pitch and translation, but we don't expect those to affect the range very much, and we see that
• Finally, we reoptimize the extrinsics, and compute a better yaw correction to "fix" the range: 0.023 degrees. This is different from the previous value of 0.03 degrees because that computation used a linearized yaw-vs-range dependence

This is all quite useful, and suggests that a small extrinsics error is likely the biggest problem.

What about --q-observation-stdev-correlation? What would be the effect of more or less correlation in our pixel observations? Running the same command with

• --q-observation-stdev-correlation 0 (the left and right pixel observations are independent) produces

## Uncertainty propagation: observation-time noise suggests worst confidence of sigma=24.06m along [ 0.084  0.329 -0.941]

• --q-observation-stdev-correlation 1 (the left and right pixel observations are perfectly coupled) produces

## Uncertainty propagation: observation-time noise suggests worst confidence of sigma=0.40m along [ 0.11   0.155 -0.982]


I.e. correlations in the pixel measurements decrease our range uncertainty. To the point where perfectly-correlated observations produce almost perfect ranging. We'll still have range errors, but they would come from other sources than slightly mismatched feature observations.

A future update to mrcal will include a method to propagate uncertainty through to re-solved extrinsics and then to triangulation. That will fill-in the biggest missing piece in the error modeling here.