7. The KITTI Benchmark Suite

Color stereo image citation

Since I have used only TSUKUBA stereo images as benchmark images, I would like to use newer benchmark images. The KITTI Vision Benchmark Suite Provides benchmark images for automated driving technology. The license is "The Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License". Here we quote two images on the left and right from a huge amount of data. Please follow this license for the reuse of these two images. Stereo Evaluation 2012 You can get it by pressing the Download stereo / optical flow data set (2 GB) button on the page.

  • quote training/colored_0/000000_10.png as L000000_10.png
  • quote training/colored_1/000000_10.png as R000000_10.png
from the file group obtained by expanding the data_stereo_flow.zip file. It seems that it is the 10th frame of the first scene 000000 of the stereo image by color. The 11th frame is also included, which seems to be an optical flow benchmark. These two images are shown below.

Application result of gaze_line-depth model

The disparity file resulting from applying depth_stereo.cpp from article 3 is

Disparity by gaze_line-depth model
. The execution was
./depth_gpu R000000_10.png L000000_10.png 68 0 14 1023 1 3
. The last 3 is the scale factor value. A maximum parallax of 68 can be found visually using 3d_stereo.cpp. The minimum parallax is 0 because a very distant view is shown. Next, the result of applying 3d_stereo.cpp is
Video generated from two stereo images
. 3d_stereo.cpp used the article 1 with the following one character key command added.
List of one key commands
6video(result.avi) out
Heye move left
Jeye move down
Keye move up
Leye move right
Feye move forward
Neye move back
Cclear angle and position
In addition, the description of the function was changed with
List of one key commands
heye move left angle
jeye move down angle
keye move up angle
leye move right angle
. Lower case "h, j, k, l, f, n keys" commands are actions around the origin, while upper case "H, J, K, L, F, N keys" commands move along the coordinate axes It becomes. Download new 3d_stereo.cpp and makefile , type
make 3d_g
, compile and start. As you can see from the makefile, the program starts with
./3d_gpu R000000_10.png L000000_10.png 68 0 14 1023 1
. To make a movie, first cut the graph with the ", key". Next, instruct the movie creation with "6 keys". In "6 keys", you can move forward, watch a little while looking forward, move forward a little, and then return to the original video. It ends with "q key". This video file result.avi is a bit large, so encode it to an m4v file using the Mac or Window function. Since it takes time to cut the graph, it is not possible with CPU.