Closed cwzat closed 4 years ago
There is no loss function for occlusion masks. The whole network is trained in an end-to-end manner after we pre-train the Flow Field Estimator. Therefore, we did not design loss functions for the mask.
How to only train the Flow Field Estimator ? I do not understand the supervisions~Thanks
You can use the following code to train the Flow Field Estimator
python train.py \
--name=fashion \
--model=pose_flow \
--attn_layer=2,3 \
--kernel_size=2=5,3=3 \
--gpu_id=0 \
--dataset_mode=fashion \
--dataroot=./dataset/fashion
We use two loss functions for the Flow Field Estimator: The Sampling Correctness Loss and The Regularization Loss. The Sampling Correctness loss is used to constrain the flow filed to sample semantically similar regions. It calculates the similarity between the warped source image and the ground truth image at the VGG feature level. The Regularization Loss is used to model the flow filed correlations of the image neighborhoods. We find that local deformations between sources and targets can be seen as the affine transformations. Therefore, we further add a regularization term to punish local regions where the transformation is not an affine transformation. Please refer to our paper for more details
Thanks, by the way,when will you release the pretrained model for shapenet?
Thanks for your interest. We are rewriting and checking the code of the ShapeNet model. I believe it will be finished before this weekend. The source code and trained model will be released once it is done.
Thanks again for your patient explaination~
Hi, in this code, the masks (output of flow net) is not under supervision. But the mask is used in
PoseTargetNet
to generate the final output. What is the loss function of mask in flow net? Or there is no loss function about masks in flow net? Thanks!