The question about pipeline for test

[Zvyozdo4ka]

Hello.

Thanks for sharing code and examples.

After reading description and paper, i thought that we give images as input to get mesh, but why in example code with bunny you loaded mesh?

# load mesh and normalize
mesh = trimesh.load_mesh('data/bunny.obj')
scale = (mesh.vertices.max(axis=0) - mesh.vertices.min(axis=0)).max()
mesh.vertices = mesh.vertices / scale * 10

I wanna get mesh of reflective object, which is assumed not to have 3d mesh yet.

Could you explain this moment? What is input? Images, intrinsic, extrinsic parameters, right?

And how do you recommend to save resulted mesh? Using trimesh or open3d?

[AtharvaJadhav7]

did you get reply @Zvyozdo4ka ??

[Zvyozdo4ka]

did you get reply @Zvyozdo4ka ??

no, i did not

[AtharvaJadhav7]

@Zvyozdo4ka I want to reconstruct less textured objects so what can I use From your github repos I can see that you have tried reconstruction of head, can you guide me on how to 3d reconstruct a hand from images

[za-cheng]

@Zvyozdo4ka Sorry for the late reply. The example script you quoted was for rendering synthetic input images from the mesh. If you already have input images you can just skip rendering and prepare inputs in the format specified there.

[za-cheng]

@AtharvaJadhav7 this repo and the method it's based on are specifically for spatially uniform BRDF and textureless (and potentially high specular) objects under a co-located camera-light source setup. Depending on your use case you might want stick to this one or try other methods which I can recommend only if I know more about your problem setting...

[Zvyozdo4ka]

@Zvyozdo4ka I want to reconstruct less textured objects so what can I use From your github repos I can see that you have tried reconstruction of head, can you guide me on how to 3d reconstruct a hand from images

hi. I tried to reconstruct hand too. For hand reconstruction, i took a video, then took frames and saved them in png format, as jpg loses some quality, then reconstructed using colmap, but as hand has too smooth texture, the reconstruction was very sparse, i think you should use sparse hand reconstruction from colmap and machine learning. You can search for some papers or github repos, who considers this problem and has models trained for your problem. Or just take colmap reconstruction and enhance it in blender

[Zvyozdo4ka]

@Zvyozdo4ka Sorry for the late reply. The example script you quoted was for rendering synthetic input images from the mesh. If you already have input images you can just skip rendering and prepare inputs in the format specified there.

Thank you for your reply, i already forgot this problem and gave up

[AtharvaJadhav7]

@AtharvaJadhav7 this repo and the method it's based on are specifically for spatially uniform BRDF and textureless (and potentially high specular) objects under a co-located camera-light source setup. Depending on your use case you might want stick to this one or try other methods which I can recommend only if I know more about your problem setting...

@za-cheng I want to reconstruct hand. I tried using SFM, MVS, but due to lack of sufficient texture and features it doesnt reconstruct properly. Also how to prepare images as input to pipeline, I can't understand from the params script, I need .npy file but not sure how to generate that. also how to get the camera poses and intrinsic and extrinsic