cvg / nice-slam

[CVPR'22] NICE-SLAM: Neural Implicit Scalable Encoding for SLAM

Home Page:https://pengsongyou.github.io/nice-slam

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

How can I get the depth l1 value as you listed in paper

pengchongH opened this issue · comments

Thanks for your nice work.

I tried to evaluate the depth l1 of Relica dataset, and I downloaded the results you provided. Then I got the following results:
accuracy: 2.941982283225373
completion: 1.8669088388589508
completion ratio: 95.0784981250763
Depth L1: 17.809303104877472

The acc and completion looks well, however, I got a bad result in depth l1.

And I also wonder which mesh you used to get the following results, I have tried final_mesh and final_mesh_rec, I got a higher completion result here.
image

I just tried to directly use the provided command line in README and the provided sample visualization's mesh to evaluate the Depth L1, and it is ~1.7cm, not 17cm for room0, maybe you can double-check.
As for Acc. and Comp. please check the difference between Table D and Table C in the paper, the code we provided is for Table C.