Warning: appropriately change of
"./utils/kitti_training_set.txt" is necessary to train from scratch the
network. To this aim, it's provided a shell script to generate a new
training file.
Warning: you can test the network simply using
"./utils/kitti_testing_set.txt" test file with images provided in
"./images" folder. If you want to predict confidence estimations with
other disparity maps use the shell script to generate a new testing
file.
The model was trained for 14 epochs, a batch size of 64, an initial
learning rate of 0.003 (reduced to 0.0003 after 10 epochs) and patches
of 9x9 extracted from 20 disparity maps computed with AD-CENSUS
algorithm on 000000_10..000019_10 stereo pairs of KITTI 12.
Results
AUC comparison between Torch and Tensorflow implementations using
disparity maps computed by AD-CENSUS algorithm on 174 testing images of KITTI 2012