search

Wuuu3511 / LAMVSNET

Boosting Multi-view Stereo with Late Cost Aggregation
11 stars 0 forks source link

code question #1

Open kk6398 opened 8 months ago

kk6398 commented 8 months ago

请问目前的代码是最终版本的代码吗？ 在networks/module.py 和 networks/mvsnet.py中没有找到论文所描述的"Cvp" 和shuffle操作 所对应的代码

YuhsiHu commented 8 months ago

Can you reproduce their results in the paper? Does late aggregation really work?

kk6398 commented 8 months ago

Can you reproduce their results in the paper? Does late aggregation really work?

I can't, I didnot find the code corresponding to the "Cvp" and shuffle operations described in the paper, and didnot find the late aggregation.

Wuuu3511 commented 8 months ago

Can you reproduce their results in the paper? Does late aggregation really work?

I can't, I didnot find the code corresponding to the "Cvp" and shuffle operations described in the paper, and didnot find the late aggregation.

Can you reproduce their results in the paper? Does late aggregation really work?

I can't, I didnot find the code corresponding to the "Cvp" and shuffle operations described in the paper, and didnot find the late aggregation.

When constructing the cost volume, we did not weight or calculate the variance of pair-wise costs (early aggregation). Instead, we preserved the pair-wise costs in the channel dimension and employed a subsequent 3D U-Net for aggregation, which we refer to as "late aggregation." To ensure that the constructed cost is independent of the input order of the source perspectives, we conducted shuffling in the channel dimension. This part of the code is implemented during data loading and you can find the corresponding code in datasets/blendedmvs.py at line 145 or datasets/dtu_yao.py at line 139.

YuhsiHu commented 8 months ago

Thank you for your prompt reply. I understand your implementation of this module. What is your baseline (overall:0.345) in DTU's ablation study?