All of my images are in RGB and in the kitti-format as recommended by TLT for training a detectnet on a custom dataset.
I used tlt-dataset-convert -d conversion_spec_file_trainval.txt -o /workspace/kitti_data/train/train_tfrecords as instructed here
The conversion_spec_file_trainval.txt is : conversion_spec_file_trainval.txt (230 Bytes)
All the training images are saved in the RGB format (544,960,3) in kitti_data/train/images/.
I see that the tfrecords I create through this command only point to the image locations.
I however do not understand why the tlt-train command decides to load the images in the nchw format instead of nhwc.