CCNN-Tensorflow

Tensorflow implementation of confidence estimation using a convolutional neural network

Learning from scratch a confidence measure
Matteo Poggi and Stefano Mattoccia
BMVC 2016

For more details:
project page
pdf

Requirements

This code was tested with Tensorflow 1.4, CUDA 8.0 and Ubuntu 16.04.

Training

Training takes about 15 minutes with the default parameters on 20 images of KITTI 2012 on a single 1080Ti GPU card.

python ./model/main.py --isTraining True --epoch 14 --batch_size 64 --patch_size 9 --dataset_training ./utils/kitti_training_set.txt --initial_learning_rate 0.003 --log_directory ./log --save_epoch_freq 2 --model_name CCNN.model

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.

./utils/kitti_generate_file.sh [path_disparities] [path_kitti_groundtruth] [index_from] [index_to] [output_file]

Testing

Test takes about 0.07 seconds on a single image of KITTI 2012  using a 1080Ti GPU card.

python ./model/main.py --isTraining False --batch_size 1 --dataset_testing ./utils/kitti_testing_set.txt --checkpoint_path ./log/CCNN.model-595140 --output_path ./output/CCNN/ad-census/

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.

./utils/kitti_generate_file.sh [path_disparities] [index_from] [index_to] [output_file]

Models

You can download a pre-trained model in ./log

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

Optimal AUC: 0.1073
Torch implementation AUC (https://vision.disi.unibo.it/~mpoggi/code.html): 0.1230
Tensorflow implementation AUC: 0.1222