Giters

ShivamDuggal4 / TARS3D

Topologically-Aware Deformation Fields for Single-View 3D Reconstruction (CVPR 2022)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ShapeNet Evaluation Results

QingkaiLu opened this issue · comments

commented

Hi Shivam,

I have ran the evaluation for the airplane and car categories of ShapeNet. The reconstructed meshes are quite impressive. I have two questions regarding the ShapeNet evaluation.

(1) For my evaluation on ShapeNet, the loss of car is 1.837e-02, the chamfer distance of car is 0.0124|0.0240. The loss of airplane is 2.289e-02, the chamfer distance of car is 0.0182|0.0160. The chamfer distances of my evaluation are smaller than that reported in the paper. For example, the chamfer distance of airplane is 0.194|0.152 in the paper. Could you please let me know what might cause the differences?

(2) I have to change the name of the ground truth point cloud from "pointcloud3.npz" to "pointcloud.npz" in this file for ShapeNet, since the name of the ShapeNet pointcloud is "pointcloud.npz" instead of ""pointcloud3.npz". Is this fine?

Thank you very much!

commented

Hi @QingkaiLu
Thanks for your warm words and your interest in our work! Please find the answers to your questions below:

  • The shared models were retrained post paper submission and after code cleaning. The slight difference in metrics could have come from the little training changes made during cleaning, for eg: by changing the number of training examples for which the additional point features were randomly turned off during stage 2 training.
  • pointcloud3.npz refers to the processed version of pointcloud.npz. To download these post-processed point clouds, please refer to this and SDF-SRN supplementary section B.1.

Hope these answer your questions!

Best Regards
Shivam

commented

Thank you for the explanation, Shivam. That is quite helpful.