Issues with Two-Stage Algorithm Performance on Clearpose Dataset

[a17843915]

I am reaching out to discuss an issue I've encountered while working with TranspareNet, specifically its application on the Clearpose dataset using the two-stage algorithm for transparent object depth completion.
My experimentation began with the default parameters provided for TranspareNet. Initially, I focused on the depth completion phase alone, and the results were reasonably good. However, when I proceeded to employ the full two-stage algorithm, encompassing both point cloud completion and depth completion, the performance significantly deteriorated.
For clarity and detailed analysis, I have attached images and results from my inference tests to this issue. These attachments illustrate the stark contrast in outcomes between the standalone depth completion and the comprehensive two-stage process.
During my troubleshooting process, I noticed that in the img2pcd file, the centering parameter is set to False. Given the context of my work and the observed performance issues, I am curious about the rationale behind this default setting. Could you please provide some insights into how this parameter influences the algorithm's performance, and whether adjusting it might improve the overall results for cases similar to mine?

Results of Point Cloud Completion：

Raw Depth：

GT Depth：