VOC predictions to KITTI

Script to create VOC style predictions to KITTI style, which is handy if you wish to run KITTI's evaluation function (e.g).

See our instructions for reproducing training and evaluation for Driving in the Matrix for our use case of this script.

It is assumed the network has been evaluated on KITTI in VOC format (e.g) so there should be 7481 images named like 000000.jpg.

VOC

VOC predictions are stored line by line, one file per detection class

<image identifier> <confidence> <left> <top> <right> <bottom>

e.g

$ head -2 path/to/output/VOC2012/Main/comp4_det_trainval_car.txt
213222 0.501 851.6 207.0 1072.3 298.9
213222 0.124 238.5 187.9 419.2 274.8

KITTI

KITTI predictions are stored line by line, one file per image, e.g

$ head -2 ~/test-kitti-container/my-kitti-labels/000000.txt
Car -1 -1 -10.000000 541.131 197.575 649.552 232.957 -1 -1 -1 -1 -1 -1 -1 0.903
Car -1 -1 -10.000000 299.747 215.436 422.328 248.200 -1 -1 -1 -1 -1 -1 -1 0.388

To run:

$ mkdir /path/to/output-labels
$ . doit.sh /path/to/voc-labels /path/to/output-labels

for example

$ mkdir /mnt/ngv/training-runs/2017-01-30-mxnet-rcnn-gta25k/evaluate-on-kitti/kitti-labels
$ . doit.sh \
  /mnt/ngv/training-runs/2017-01-30-mxnet-rcnn-gta25k/evaluate-on-kitti/VOC2012/Main \
  /mnt/ngv/training-runs/2017-01-30-mxnet-rcnn-gta25k/evaluate-on-kitti/kitti-labels