Updated KITTI devkit Code for Error Maps, Color Disparity Generation, and pfm2uint16png and uint16png2pfm Converting

Compilation

The CMakeLists.txt file is provided for convenient compilation. Just run the following:

cd kitti-devkit # assuming its the root dir for this reporsitory
mkdir build
cd build
cmake ..
make

The compiled libraries or executables will be found at ./lib and ./build sub-directory, respectively.

Change your pfm disparity to png format:

Assuming your disparity prediciton is in .pfm format. KITTI devkit code accepts PNG disparity maps for evaluation. So firstly, coverting the pfm to png ones.

# line 44 in file kt_errorMap.sh
./build/pfmDisp2png $pfm_file $png_file

where, executable pfmDisp2png will convert source file $pfm_file to target file $png_file. The detailed commands are shown in file kt_errorMap.sh.

Evaluation for Error Maps and Color Disparities

Based on the png disparity maps, we can do evaluation, via:

KITTI 2015:

# kitti 2015: line 48 in file kt_errorMap.sh
./build/evaluate_stereokt15 $result_sha $isLogColor

where, $result_sha is the disparity input for evaluation, and $isLogColor is a flag for using KITTI LogColor (if isLogColor=true) or not (if isLogColor=false).

This will result in

the error maps at errors_disp_img_0/,
the color disparity maps at result_disp_img_0/,

both under the same location as your input disparity maps for evaluation.

KITTI 2012:

# kitti 2012: line 77 in file kt_errorMap.sh
./build/evaluate_stereokt12 $result_sha $isLogColor

where, $result_sha is the disparity input for evaluation, and $isLogColor is a flag for using KITTI LogColor (if isLogColor=true) or not (if isLogColor=false).

This will result in

the error maps at errors_img/,
the color disparity maps at disp_ipol/ and disp_orig/,

both under the same location as your input disparity maps for evaluation.

One Example from KITTI 2015:

The following shows the input disparity, the color version disparity, error map (w/o and w/ log color), and the KITTI error scale:

Generate KITTI-color-map Disparities for Middlebury 2014, ETH3D (and KITTI of course)

The executable ./build/get_color_disp is used to directly convert your .pfm disparity prediction to color .png files.

Conver disparity prediction to color png version

We provide the bash file get_color_disp.sh for Middlebury 2014 and KITTI 2015.

#line 44
app=./build/get_color_disp
# for middlebury, line 57
$app $gt_disp_file $input_disp_file $output_disp_file

# for KITTI 2015, line 84
$app $gt_disp_file $input_disp_file $output_disp_file

where,

$app is the executable;
$gt_disp_file is the ground truth disparity;
$input_disp_file is the input disparity, i.e., your predicted disparity file in pfm format;
$output_disp_file is the color png disparity output.

Convert ground truth disparity to interpolated color png version

Still use the exectuable get_color_disp, with $input_disp_file set be equal to $gt_disp_file. See the bash file get_color_disp.sh for details.

Take a hard image 000104_10.png from KITTI 2015 as an example, we show the reference image, color interpolated ground truth disparity, color disparity predition, and error map and its error scale.

A Simple Python Code for bad-x (e.g, x = 3.0) error calculation

We provide a simple python code, shown in the file kt_bad3_error.py, to calculate bad-x (e.g, x=3.0) error for each input image. See this file for details.

ccj5351/kitti-devkit