search

KaiqiangXiong / CL-MVSNet

[ICCV2023] CL-MVSNet: Unsupervised Multi-view Stereo with Dual-level Contrastive Learning
MIT License
40 stars 3 forks source link

ICCloss and image-level data augmentation #10

Open CUICHENXI222 opened 2 weeks ago

CUICHENXI222 commented 2 weeks ago

Hello, while reviewing your code, I noticed that the mask used for calculating Iccloss is a randomly generated filter_mask. Additionally, during the image-level data augmentation process, you not only add Bernoulli noise but also apply the randomly generated filter_mask to the reference image again. This seems inconsistent with the description in the paper. What is the reasoning behind this approach? Below is the relevant source code:

ref_img, filter_mask = random_image_mask(ref_img, filter_size=(ref_img.size(2) // 3, ref_img.size(3) // 3))

KaiqiangXiong commented 1 week ago

It is the Cross-view Masking technique proposed in JDACS(Self-supervised Multi-view Stereo via Effective Co-Segmentation and Data-Augmentation), which is supposed to be mentioned along with "the color fluctuation augmentation" in the last sentence "Note that the color fluctuation augmentation used by [5, 41, 42] is also applied in the image-level contrastive branch" of Section3.5 of the paper. More details can be found in JDACS.