search

VCL3D / Pano3D

Code and models for "Pano3D: A Holistic Benchmark and a Solid Baseline for 360 Depth Estimation", OmniCV Workshop @ CVPR21.
https://vcl3d.github.io/Pano3D
MIT License
82 stars 7 forks source link

How to understand the mean of the delta accuracyand completeness? #3

Closed Liusandian closed 2 years ago

Liusandian commented 2 years ago

we noticed that when compute the metric of mask boundry,the code in pytorch is "D_gt = ndimage.distance_transform_edt(1-gt.cpu())",how to understand the distance transform? And why we should use the 1-gt_edge as the input of compute the distance transform?

Looking forward your reply,thanks!

Liusandian

tzole1155 commented 2 years ago

Hi Liusandian,

As mentioned in our paper, and also in the code, the boundary metrics and their implementation are based on the work proposed by Tobias Koch (https://openaccess.thecvf.com/content_ECCVW_2018/papers/11131/Koch_Evaluation_of_CNN-based_Single-Image_Depth_Estimation_Methods_ECCVW_2018_paper.pdf). There you can find more details and visual examples regarding these metrics.

As we are focusing on the closest distance to the boundary of each object for each pixel, and not to the background, we need to apply the image distance transform in the inverted edge image.

Hope this helps!

Best,

Giorgos