dispflownet-tf

Reimplementation in tensorflow of Dispnet and Dispnet-Corr1D. Original code thx to @fedor-chervinskii

Improvement and fixes

• Python3 conversion
• Add compile.sh inside user_ops
• Modified image preprocessing to match those used in caffe
• Added implementation of "Unsupervised Adaptation for Deep Stereo" - ICCV2017 - from now on shortened 'UA'
• Lots of arguments from command line
• Native tensorflow input pipeline
• Add improved code for inference

Pretrained nets

• Weights for Dispnet with tensorflow correlation layer trained for 1200000 step on FlyingThings3D available here

Training

1. Create a training and validation set made of couple of left and right frames + disparities (+ confidence for UA)
2. Create a txt file with the list of training samples for your trainign set, each row on the file should contain "path_left_frame;peth_right_frame;path_ground_truth", for UA "path_left_frame;peth_right_frame;path_disparity;path_confidence". For UA "path_disparity" is the path to the disparity map obtained by a standard, non learned, stereo algorithm (e.g.: SGM or AD-CENSUS in the paper)
3. Train Dispnet using train.py, usage:

python main.py --training $TRAINING_LIST --testing $TEST_LIST -c $OUT_DIR --corr_type tf 

Arguments, some are optional or already provide a default value:

• --training TRAININGLIST: path to the training list as defined at 2.
• --testing TEST_LIST: path to the test list as defined at 2.
• -c CHECKPOINT_PATH: path were the log and trained CNN will be saved.
• -b BATCH_SIZE: number of samples for each train iteration.
• -l LOG_STEP: every how many step save summaries.
• -w FILE: optional initialization weights.
• -s SAVE_STEP: every how many step save the current network.
• -n N_STEP: number of training step to perform
• --corr_type [tf,cuda,none]: type of correlation layer implemented in dispnet ('tf'->pure tensorflow implementation, 'cuda'->native cuda implementation, needs to manually compile the op defined in user_ops/, 'none'->no correlation layer)
• --kittigt: load gt map as 16bit png image with each pixel encoding=disparity x 256

Additional arguments for UA:

• -th CONFIDENCE_TH: minimum confidence value to consider a gt pixel as valid
• --smooth SMOOTH: multiplier for the smoothing twerm of the loss function
• --doubleConf: flag to read confidence map as 16bit png image with each pixel encoding=confidence x 256 x 256

Test/inference

inference.py can be used to perform inference with any kind of Dispnet, even the one trained using UA, on a list of stereo frames and save the resulting disparities on disk.

python inference.py --left $LEFT_FOLDER --right $RIGHT_FOLDER --ckpt $CKPT_PATH -o $OUT_FOLDER --fullRes -v

Arguments, some are optional or already provide a default value:

• --left LEFT: either path to a folder containing the left frames or to a txt file with the list of left frame paths to load.
• --right RIGHT: either path to a folder containing the right frames or to a txt file with the list of left frame paths to load.
• --ckpt CKPT: path to the weight of Dispnet that needs to be loaded.
• --corr_type [tf,cuda,none]: type of correlation layer implemented in dispnet ('tf'->pure tensorflow implementation, 'cuda'->native cuda implementation, needs to manually compile the op defined in user_ops/, 'none'->no correlation layer)
• -o OUT_DIR: path to the output dir where the results will be saved
• -v: flag to enable visualization
• --fullRes: flag to save the output of the network rescaled at the full input resolution
• --max_disp MAX: maximum value for disparity, value above will be clipped before saving the disparities.