OpFlowCNN

Optical flow estimation using convolutional neural network

Run training

Dataset for training can be found in data/synt-mat Custom dataset can be created using matlab script dataset_m.m

Arguments for training

• CUDA_VISIBLE_DEVICES - selecting gpu to run training on
• data_dir - where is placed the dataset to be used for training and validation
• result_dir - directory for the results of the training
• sgd - decides if training will use sgd or adadelta (1 = sgd)
• model - select from models (1 = model-small, 2 = model-basic, 3 = model-parallel)
• batchSize - size of the batch
• print_freq - how often will the results be saved
• print_img - should the results be saved
• save_model - should the model be saved (will be saved in models-learned)
• dataSize - size of the training set
• valSize - size of the validation set

Train using OpticalFlowCriterion

CUDA_VISIBLE_DEVICES=0 th train-gpu.lua -data_dir data/ -result_dir 10_06 -sgd 1 -model 2 -batchSize 10 -epochs 3000 -print_freq 100 -print_img 1 -save_model 0 -dataSize 900 -valSize 100

Train using MSECriterion (Mean square error)

CUDA_VISIBLE_DEVICES=0 th train-mse.lua -data_dir data/ -result_dir 10_06-mse -sgd 1 -model 2 -batchSize 1 -epochs 5000 -print_freq 100 -print_img 1 -save_model 1 -dataSize 900 -valSize 100