Method

A Self-Supervised Permutation Approach to the Stereo Matching Problem [Permutation Stereo]


Submitted on 5 Mar. 2022 20:33 by
Pierre-Andre Brousseau (Universite de Montreal)

Running time:30 s
Environment:GPU @ 2.5 Ghz (Matlab)

Method Description:
This paper proposes a novel permutation
formulation to the stereo matching problem. Our
proposed approach introduces a permutation volume
which provides a natural representation of stereo
constraints and disentangles stereo matching from
monocular disparity estimation. It also has the
benefit of simultaneously computing disparity and
a confidence measure which provides explainability
and a simple confidence heuristic for occlusions.
In the context of self-supervised learning, the
stereo performance is validated for standard
testing datasets and the confidence maps are
validated through stereo-visibility. Results show
that the permutation volume increases stereo
performance and features good generalization
behaviour. We believe that measuring confidence is
a key part of explainability which is instrumental
to adoption of deep methods in critical stereo
applications such as autonomous navigation.
Parameters:
Our model is trained on the datasets at half
resolution on random image crops of size of 192 ×
32 pixels with a batch size of 2. No other data
augmentation is applied apart from random crops.
The conv blocks are as defined in [18] and their
Fig. 5 (right) with f=32. The convolution layers
apply fixed padding, have batch normalization and
have a Relu activation function. The
implementation is made with Mathematica[21] 12.3.
The λ is set to 10 and the symmetric normalization
has t=8 iterations. Networks are trained until
convergence with the Adam Optimizer[24] and a
learning rate of 1×10−3. The constant α is set to
0.85 as is customary [7] and τ is set to 0.1. The
models are trained on an RTX3090.
Latex Bibtex:
@inproceedings{brousseau2022permutation,
title={A Permutation Model for the Self-
Supervised Stereo Matching Problem},
author={Brousseau, Pierre-Andr{\'e} and Roy,
S{\'e}bastien},
booktitle={2022 19th Conference on Robots and
Vision (CRV)},
pages={122--131},
year={2022},
organization={IEEE}
}

Detailed Results

This page provides detailed results for the method(s) selected. For each of the first 20 test images, the number of erroneous pixels at all thresholds is depicted in the table. Underneath, the left input image, the disparity / end-point error map and the estimated (and interpolated) disparity / optical flow map are shown. The error map scales linearly between 0 (black) and >=5 (white) pixels error. Red denotes all occluded pixels, falling outside the image boundaries. The false color map is scaled to the largest ground truth disparity / flow value.

Test Set Average

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 11.89 % 13.16 % 1.6 px 1.8 px
3 pixels 7.39 % 8.48 % 1.6 px 1.8 px
4 pixels 5.41 % 6.34 % 1.6 px 1.8 px
5 pixels 4.32 % 5.11 % 1.6 px 1.8 px
This table as LaTeX

Reflective Regions

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 35.75 % 38.59 % 5.0 px 5.8 px
3 pixels 26.86 % 29.87 % 5.0 px 5.8 px
4 pixels 21.96 % 24.90 % 5.0 px 5.8 px
5 pixels 18.78 % 21.60 % 5.0 px 5.8 px
This table as LaTeX

Test Image 0

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 9.36 % 11.12 % 1.2 px 1.4 px
3 pixels 6.62 % 8.16 % 1.2 px 1.4 px
4 pixels 4.97 % 6.29 % 1.2 px 1.4 px
5 pixels 3.94 % 4.88 % 1.2 px 1.4 px
This table as LaTeX





Test Image 1

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 12.61 % 14.72 % 1.6 px 1.7 px
3 pixels 7.16 % 8.84 % 1.6 px 1.7 px
4 pixels 4.83 % 6.14 % 1.6 px 1.7 px
5 pixels 3.46 % 4.40 % 1.6 px 1.7 px
This table as LaTeX





Test Image 2

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 11.87 % 12.61 % 1.2 px 1.2 px
3 pixels 6.98 % 7.44 % 1.2 px 1.2 px
4 pixels 4.49 % 4.75 % 1.2 px 1.2 px
5 pixels 3.15 % 3.30 % 1.2 px 1.2 px
This table as LaTeX





Test Image 3

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 23.58 % 25.35 % 2.1 px 2.3 px
3 pixels 16.00 % 17.71 % 2.1 px 2.3 px
4 pixels 12.10 % 13.65 % 2.1 px 2.3 px
5 pixels 9.90 % 11.30 % 2.1 px 2.3 px
This table as LaTeX





Test Image 4

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 12.58 % 13.51 % 1.1 px 1.2 px
3 pixels 3.71 % 3.90 % 1.1 px 1.2 px
4 pixels 1.37 % 1.60 % 1.1 px 1.2 px
5 pixels 0.76 % 0.96 % 1.1 px 1.2 px
This table as LaTeX





Test Image 5

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 4.78 % 4.96 % 1.1 px 1.1 px
3 pixels 2.45 % 2.55 % 1.1 px 1.1 px
4 pixels 1.70 % 1.78 % 1.1 px 1.1 px
5 pixels 1.45 % 1.49 % 1.1 px 1.1 px
This table as LaTeX





Test Image 6

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 13.30 % 14.80 % 1.9 px 2.1 px
3 pixels 8.04 % 9.14 % 1.9 px 2.1 px
4 pixels 6.16 % 7.12 % 1.9 px 2.1 px
5 pixels 5.17 % 5.78 % 1.9 px 2.1 px
This table as LaTeX





Test Image 7

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 9.87 % 12.04 % 1.6 px 1.9 px
3 pixels 6.03 % 8.02 % 1.6 px 1.9 px
4 pixels 4.43 % 6.14 % 1.6 px 1.9 px
5 pixels 3.62 % 4.98 % 1.6 px 1.9 px
This table as LaTeX





Test Image 8

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 5.05 % 5.67 % 0.9 px 1.0 px
3 pixels 2.17 % 2.72 % 0.9 px 1.0 px
4 pixels 1.60 % 2.15 % 0.9 px 1.0 px
5 pixels 1.33 % 1.76 % 0.9 px 1.0 px
This table as LaTeX





Test Image 9

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 15.93 % 18.01 % 1.9 px 2.2 px
3 pixels 9.50 % 11.63 % 1.9 px 2.2 px
4 pixels 6.11 % 8.29 % 1.9 px 2.2 px
5 pixels 4.88 % 6.67 % 1.9 px 2.2 px
This table as LaTeX





Test Image 10

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 7.65 % 8.83 % 1.0 px 1.1 px
3 pixels 3.87 % 4.81 % 1.0 px 1.1 px
4 pixels 2.31 % 3.09 % 1.0 px 1.1 px
5 pixels 1.65 % 2.38 % 1.0 px 1.1 px
This table as LaTeX





Test Image 11

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 17.90 % 20.42 % 2.0 px 2.2 px
3 pixels 10.02 % 12.38 % 2.0 px 2.2 px
4 pixels 7.08 % 9.08 % 2.0 px 2.2 px
5 pixels 5.87 % 7.38 % 2.0 px 2.2 px
This table as LaTeX





Test Image 12

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 12.70 % 14.94 % 2.2 px 3.9 px
3 pixels 9.08 % 11.40 % 2.2 px 3.9 px
4 pixels 7.53 % 9.89 % 2.2 px 3.9 px
5 pixels 6.53 % 8.91 % 2.2 px 3.9 px
This table as LaTeX





Test Image 13

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 13.82 % 14.96 % 2.0 px 2.2 px
3 pixels 9.78 % 10.85 % 2.0 px 2.2 px
4 pixels 7.75 % 8.74 % 2.0 px 2.2 px
5 pixels 6.60 % 7.45 % 2.0 px 2.2 px
This table as LaTeX





Test Image 14

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 22.36 % 22.81 % 1.8 px 1.8 px
3 pixels 14.40 % 14.60 % 1.8 px 1.8 px
4 pixels 9.78 % 9.61 % 1.8 px 1.8 px
5 pixels 7.61 % 7.44 % 1.8 px 1.8 px
This table as LaTeX





Test Image 15

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 16.09 % 16.31 % 1.7 px 1.7 px
3 pixels 10.25 % 10.38 % 1.7 px 1.7 px
4 pixels 6.57 % 6.62 % 1.7 px 1.7 px
5 pixels 4.94 % 4.89 % 1.7 px 1.7 px
This table as LaTeX





Test Image 16

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 5.61 % 5.77 % 0.9 px 0.9 px
3 pixels 2.49 % 2.47 % 0.9 px 0.9 px
4 pixels 1.35 % 1.32 % 0.9 px 0.9 px
5 pixels 0.93 % 0.91 % 0.9 px 0.9 px
This table as LaTeX





Test Image 17

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 10.84 % 11.92 % 1.4 px 1.5 px
3 pixels 7.87 % 8.81 % 1.4 px 1.5 px
4 pixels 6.33 % 7.01 % 1.4 px 1.5 px
5 pixels 5.17 % 5.50 % 1.4 px 1.5 px
This table as LaTeX





Test Image 18

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 9.89 % 10.59 % 1.1 px 1.1 px
3 pixels 6.72 % 7.26 % 1.1 px 1.1 px
4 pixels 4.97 % 5.47 % 1.1 px 1.1 px
5 pixels 3.71 % 4.13 % 1.1 px 1.1 px
This table as LaTeX





Test Image 19

Error Out-Noc Out-All Avg-Noc Avg-All
2 pixels 5.08 % 5.48 % 1.0 px 1.0 px
3 pixels 2.36 % 2.33 % 1.0 px 1.0 px
4 pixels 1.57 % 1.53 % 1.0 px 1.0 px
5 pixels 1.19 % 1.15 % 1.0 px 1.0 px
This table as LaTeX







eXTReMe Tracker