search

yuxiaoguo / VVNet

Implementation of View-volume network for semantic scene completion from a single depth image
MIT License
15 stars 3 forks source link

understanding feature projection and .mat file #9

Closed NguyenTriTrinh closed 4 years ago

NguyenTriTrinh commented 4 years ago

Hi, In the 3.2 feature projection part ,compared with SSCNet ,you replaced the TSDF with feature projection right? when depth map passed by 2D view network,there was a pooling operation, after that, several neighboring features were projected into the same voxel in the projection , so there were two pooling operations in total? You used the feature maps constructed by the 2D view network as the input to the projection instead of the original depth map, am i right? Another question,when i load the XXXX_gt_d4.matand XXXX_vol_d4.mat, i get distance_ds,flipVol_ds,sceneVox_ds,what are they respectively mean? it seems to me that distance_dsrepresents the tsdf values, and flipVol_dsrepresents the ftsdf values.

yuxiaoguo commented 4 years ago

Sorry for the late response.

About the feature projection, your understanding is totally right. During the projection, many pixels may fall into an identity voxel. Thus, an average pooling operation is imported to handle this case.

About the content in MAT files, I don't fully understand the content as well. Fortunately, we just need to derive the surface and non-surface information. For this propose, you may refer to the analysis/statistics.py. For other proposes, ask the SSCNet author may a better choice.

NguyenTriTrinh commented 4 years ago

thank u for your reply! i have already understood.

------------------ 原始邮件 ------------------ 发件人: "yuxiaoguo"<notifications@github.com>; 发送时间: 2019年12月5日(星期四) 下午2:23 收件人: "yuxiaoguo/VVNet"<VVNet@noreply.github.com>; 抄送: "雤 ﹊踜|〆.."<644538318@qq.com>; "State change"<state_change@noreply.github.com>; 主题: Re: [yuxiaoguo/VVNet] understanding feature projection and .mat file (#9)

Sorry for the late response.

About the feature projection, your understanding is totally right. During the projection, many pixels may fall into an identity voxel. Thus, an average pooling operation is imported to handle this case.

About the content in MAT files, I don't fully understand the content as well. Fortunately, we just need to derive the surface and non-surface information. For this propose, you may refer to the analysis/statistics.py. For other proposes, ask the SSCNet author may a better choice.

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub, or unsubscribe.