Labels within camera 00

[ZiadElmassik]

Hello @mpitropov, I deeply apologize for the spam. There's a problem with the 3d_annotations.json file in which all the labels within all the cameras are included. I'm trying to only convert the labels within camera 00 to KITTI format to evaluate the models. Are there any identifiers within the 3d_ann.json file or any way to identify which labels reside within which camera view?

[mpitropov]

Hi, in some of my example demo files I did the following:

1. Transform the cuboid to the camera frame that I want, followed by filtering out negative z values
2. Transforming the individual cuboid corners to the image frame and filtering out negative z values
3. Divide x,y values by z to get the image coordinates and filter out cuboids with values outside of the image.

[ZiadElmassik]

Hello @mpitropov, I've nearly perfected the conversion script. All that remains is to simply filter out bounding boxes that are outside camera 00's FOV. Can you please specify which demo contains the filtering?

[mpitropov]

This script: https://github.com/mpitropov/cadc_devkit/blob/master/run_demo_tracklets.py

cadc_devkit/run_demo_tracklets.py

Line 72 in 9c43a31

if tmp[2][3] < 0: # Behind camera

2. Multiple of these lines:
   

   cadc_devkit/run_demo_tracklets.py

   Line 91 in 9c43a31

   if tmp[2][3] < 0:

3. Multiple of these lines:
   

   cadc_devkit/run_demo_tracklets.py

   Line 130 in 9c43a31

   if f_r_b_coord[0] < 0 or f_r_b_coord[0] > img_w or f_r_b_coord[1] < 0 or f_r_b_coord[1] > img_h:

The filtering could be modified. Currently, it removes cars if even one corner point is outside of the image coordinates.