train with custom data

[wmrenr]

Hello, first of all, thank you for your great contribution! I trained the network with custom 60 images, and the reconstructed model contains holes, as shown in the following figure. I tried to adjust the parameter training, but still couldn't solve the problem. Can you help me？ And I found that some of the reconstruction results I trained with images of other objects didn't contain holes. So I guess this training result is related to the collection effect of images, so how can images be collected be beneficial for the experimental results? Can you give some suggestions?

[wutong16]

Hi!

I believe the hole could potentially be attributed to an inaccurate bounding box estimation during the data preprocessing phase. Please examine whether the hole resembles a typical cut created by a plane.

To address this issue, you can begin by manually increasing the bounding box area by adding a small buffer (e.g., 0.1) to xyz_max and subtracting a small buffer from xyz_min.

Additionally, when capturing the video, strive to keep the object approximately centered within the frames and maintain a clean background. These practices will facilitate the production of precise masks and improve the accuracy of the bounding box estimation process.

[wmrenr]

Thank you for your guidance! I observed the hole and found that it looks like a typical cut created by a plane. But I don't understand where to add a small buffer (e.g., 0.1) to xyz_max and subtracting a small buffer from xyz_min. Are they the compute_bbox_by_cam_frustrm() function and the compute_bbox_by_coarse_geo() function in the run.py file？

[wutong16]

Sorry for the delayed follow-up! Please change the setting here, where the model is being initialized.