Home

Awesome

Unsupervised-Confidence-Measures

Learning confidence measures in the wild
Fabio Tosi, Matteo Poggi, Alessio Tonioni, Luigi Di Stefano and Stefano Mattoccia
BMVC (British Machine Vision Conference) 2017

This strategy provides labels for training confidence measures based on machine-learning technique without ground-truth labels. Compared to state-of-the-art, this method is neither constrained to image sequences nor to image content.

Warning: This C++ implementation contains AD-CENSUS stereo matching algorithm only.

KITTI 2012 sample

Alt text Alt text Alt text

Requirements

This C++ code is developed under Ubuntu 16.04. The following libraries are required:

Building from the command line

Create a build directory and run cmake there:

Usage

Input (required):

-l <left_image> : left image of stereo pair 
-r <right_image> : right image of stereo pair
-o <output_path> : output folder  
-p <positive_cms> : list of confidence measures used to select positive samples
-n <negative_cms> : list of confidence measures used to select negative samples

Input (optional):

-d <d_max> : max disparity range (default=228)
-t <threshold> : percentage of pixels considered most/least confident (default=0.4)
-b <bad> : value used to discriminate positive from negative samples (default=3)
-g <groundtruth> : if groundtruth is available you can test the accuracy of positive and negative samples
-f <evaluation_file> : CSV file in which the evaluation will be saved (if -g arg is enabled)
-i <invalid_pixel> : invalid pixel value for gt image (e.g 0 for KITTI, 255 for Middlebury) 
-s <scale_factor> : scale factor for gt image (e.g 256 for KITTI, 1 for Middlebury)

Example without groundtruth for test:

./build/bmvc2017 -l [left_image] -r [right_image] -o [output_path] -p lrc uc dbl apkr med wmn -n lrc uc apkr wmn -t 0.3 -b 3 -d 228

Example with groundtruth for test:

./build/bmvc2017 -l [left_image] -r [right_image] -g [gt_image] -f [csv_file] -o [output_path] -p lrc uc dbl apkr med wmn -n lrc uc apkr wmn -t 0.3 -b 3 -d 228 -i 0 -s 256

List of usable confidence measures (CMs)

 - msm (Matching Score Measure)
 - lc (Local Curve)
 - cur (Curvature)
 - pkr (Peak Ratio)
 - pkrn (Peak Ratio Naive)
 - apkr (Average Peak Ratio)
 - apkrn (Average Peak Ratio Naive)
 - mm (Maximum Margin)
 - mmn (Maximum Margin Naive)
 - nmm (Nonlinear Maximum Margin)
 - nmmn (Nonlinear Maximum Margin Naive)
 - dvm (Disparity Variance Measure)
 - dam (Disparity Ambiguity Measure)
 - mlm (Maximum Likelihood Measure)
 - aml (Attainable Maximum Likelihood)
 - nem (Negative Entropy Measure)
 - noi (Number Of Inflections)
 - wmn (Winner Margin)
 - wmnn (Winner Margin Naive)
 - per (Perturbation Measure)
 - lrc (Left Right Consistency Check)
 - acc (Asymmetric Consistency Check)
 - uc (Uniqueness Constraint)
 - lrc (Left Right Difference)
 - hgm (Horizontal Gradient Measure)
 - zsad (Zero Mean Sum Of Absolute Difference)
 - dsm (Distinctive Similarity Measure)
 - samm (Self Aware Matching Measure)
 - dbl (Distance To Border Left)
 - dtd (Distance To Discontinuity)
 - dte (Distance To Edge)
 - med (Difference With Median)
 - lmn (Local Minima In Neighborhood)
 - da (Disparity Agreement)
 - ds (Disparity Scattering)

For more details about confidence measures: Quantitative evaluation of confidence measures in a machine learning world

Reference

If you use this code, please cite our paper:

@inproceedings{BMVC_2017,
   author               = {Tosi, Fabio and Poggi, Matteo and Tonioni, Alessio and Di Stefano, Luigi and Mattoccia, Stefano},
   booktitle            = {28th British Machine Vision Conference (BMVC 2017)},
   month                = {September},
   title                = {Learning confidence measures in the wild},
   year                 = {2017},
   }