Home

Awesome

dispflownet-tf

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

Improvement and fixes

Pretrained nets

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:

Additional arguments for UA:

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: