KITTI Data Trains in TF but not in Digits

[DennisFaucher]

Hi.

I am having a devil of a time labeling object recognition data so that DIGITS will train it. My goal is to move a caffe model to my Xavier. I decided to use someone else's successful KITTI data as a test instead. I this blog, KITTI-labeled data is successfully trained in tensorflow:

https://jkjung-avt.github.io/hand-detection-tutorial/

When I take that data, split in to train and val folders (4,320 images & 480 images) and train in DIGITS using these instructions:

https://github.com/NVIDIA/DIGITS/tree/master/examples/object-detection

MAP and Precision never rise above zero.

Here are labels for comparison:

Blog Hand KITTI Data Label hand 0 0 0 2 446 314 644 0 0 0 0 0 0 0 0 hand 0 0 0 555 540 1104 717 0 0 0 0 0 0 0 0 hand 0 0 0 539 245 829 560 0 0 0 0 0 0 0 0 hand 0 0 0 2 338 76 469 0 0 0 0 0 0 0 0

NVIDIA Example KITTI Label Car 0.00 0 -1.84 662.20 185.85 690.21 205.03 1.48 1.36 3.51 5.35 2.56 58.84 -1.75 Van 0.00 0 1.70 448.07 177.14 481.60 206.41 2.50 2.20 5.78 -13.02 2.91 65.02 1.50 DontCare -1 -1 -10 610.50 179.95 629.68 196.31 -1 -1 -1 -1000 -1000 -1000 -10 DontCare -1 -1 -10 582.97 182.70 594.78 191.05 -1 -1 -1 -1000 -1000 -1000 -10 DontCare -1 -1 -10 600.36 185.59 608.36 192.69 -1 -1 -1 -1000 -1000 -1000 -10

I have read that fields 1,5,6,7,8 are all that are needed for object detection training.

Any ideas? TIA.