the mesh of my own data is bad

[Gary-Qinghua]

the mesh of my own data is bad

[ashawkey]

@Gary-Qinghua Hi, I guess you'll want to try the SDF mode for single-object reconstruction. The scale and bound may need to be adjusted. Also, many photos are sufferring from motion-blur. You may follow the data capture advice here.

[Gary-Qinghua]

Thanks. I tried SDF mode, but it seems less effective. What you said should be right. I will collect again according to your advice.
I have another question about the mesh quality. Due to the large number of spikes, the result of stage0 looks better than stage1. Why?
Can I not use this function "refine_and_decimate " in stage1?

[Gary-Qinghua]

ngp_stage0_ep0286_rgb.mp4

ngp_stage1_ep0096_rgb.mp4

[ashawkey]

Yes, and that's what SDF mode will do. I guess you should apply background removal for SDF mode.

[Gary-Qinghua]

If you just reconstruct a single object, a lot of tasks can be done. I think getting the mesh of scene is the highlight of your project. The results on the data of MIP-NERF-360 are very good.

[ashawkey]

In that case you may try to set opt.refine = False manually, or increase --lambda_normal to a very high value like 0.1. Actually SDF mode can somewhat work for scene reconstruction now, you could also give it a try. If the model won't converge, you can try to turn on '--enable_dense_depth'.

[Gary-Qinghua]

I will try. Thanks!