Conversion from 3-D to 2-D annotations results in empty annotation files

[kkannan8291]

@mpitropov I using the script 'cadc_unpack_all_kitti.py' in order to convert the 3-D annotations to 2-D annotations and make it compliant similar to the Kitti dataset. While this seems to work for 55% of the CADC dataset, the 2-D annotations (.txt files generated) are empty for the remaining images. I am seeing this issue across all cameras. While the number of images without annotations are less than 50 for Cam00, 01 and 07, the other cameras have over 50% of the dataset without any labels. Especially for Cam04, the entire dataset has no annotations.

On closer introspection into the 3-D annotations.json file, I noticed that the 'camera used' parameter is set to 'null' for over 90% of the dataset (Let's take scenario 2 [2018_03_06] as an example). Out of the 5185 usages of 'camera_used', 4398 of them are set to 'null'. Can you please explain what this exactly means?

Do you think this could be the root cause for the empty 2-D annotation files that are being generated?

reference: https://github.com/mathild7/faster_rcnn_pytorch_multimodal/blob/3145d91cba5ec38769b76af8f6e118748876c280/tools/cadc_unpack_all_kitti.py

[mpitropov]

Scale AI were the ones who labelled the dataset and the camera used parameter is filled in if they used the camera during labelling. I think it will be null if they didn't use the camera images and only relied on lidar.

There is also an updated converter script in this issue you can try https://github.com/mpitropov/cadc_devkit/issues/21