The result
Closed this issue · 2 comments
I have a question to ask. In the paper, the D1-all value for KITTI 2015 is 2.03. However, in the KITTI 2015 stereo benchmark, the D1-all value is 1.59. Was this improvement achieved by adjusting parameters or implementing other effective methods?
Hi,
In the ICCV'23 camera-ready version (which corresponds to version v3 on arxiv) we indeed report 1.59 D1-all, as you see in the official kitti 15 leaderboard.
The 2.03 D1-all you mentioned was from the version v2 of arxiv from March 2023.
The main difference between the two versions is that we finetune on higher-resolution crops (1216x352) instead of 704x352. There are also some minor differences that should have low impact, i.e. we bound the confidence in the loss, etc.
Best
Philippe
Thanks for your response!