Training and Testing issues about intrinsic_flow HOT 1 CLOSED

Hi SandhyaLaxmi. Thanks for being interested in our work.

For the Q1, you can use it like python ./scripts/flow_regression_module.py --gpu_ids 0 --id model_id. Also, train_flow_regression_module.py will perform testing and output testing information after each training epoch.

For Q2, I think other segmentation data will also help. The segmentation information is used in the training of the flow regression module to penalize pixels warped to incorrect regions, which we think is like hard negative mining. The reason we use SMPL to generate segmentation data instead of using other segmentation model/dataset is to avoid involving extra knowledge.

For Q3, you can test pose_transfer_0.1 just like other model by python scripts/test_pose_transfer.py --gpu_ids 0 --id PoseTransfer_0.1 --which_epoch best --save_output. Please note that pose_transfer_0.1 has a normal U-Net architecture without a dual encoder.

from intrinsic_flow.