Poor Performance on ScanNet

[wen-yuan-zhang]

Thanks for you excellent work! I have tried it on ScanNet datsets, however, I got a totally bad results. This is the reconstruction results after 30k iters:

where the gt mesh is like:

I haved adjusted the dataset format and config files, but I don't know whether the result is due to the method or due to some hyperparameters. The adjusted config file is below:

model:
    object:
        sdf:
            mlp:
                inside_out: True
                out_bias: 0.6
            encoding:
                coarse2fine:
                    init_active_level: 8
    appear_embed:
        enabled: True
        dim: 8
    background:
        enabled: False
    render:
        num_samples:
            background: 0

Additinally, I normalize ScanNet scene into the cube of [-0.5, 0.5], and I changed the near/far sampling to 0 and 1.5

Do you have any ideas why ScanNet appears so poor performance?

[chris-aeviator]

Given what I understand from neuralangelo, more iterations = finer details, the default setting for iterations are 500k in the config, might it be you were not training long enough?

(I will have a comparasion of 50k steps vs. 500k steps in about 1 hr.)

[chenhsuanlin]

Hi @zParquet, it could very well be possible that it has not been trained long enough yet. Do the results in your visualizer (e.g. W&B) look reasonable? Additionally, I've pushed a small fix on the mesh extraction script (#41), which may be related to your issue. If you could pull again and still see the same issue, please let me know.

[chris-aeviator]

50k steps have been a mere blob whereas 500k steps are usable (this was bad input material btw.)

50k steps, hashgrid.dict_size 10

500k steps, hashgrid.dict_size 16

[chenhsuanlin]

Nice results, thanks @chris-aeviator for sharing! Have you tried using a larger dict_size value (default was 22)? Would also be great if you could try extracting the mesh with textures (#45)!

[MulinYu]

Thanks for you excellent work! I have tried it on ScanNet datsets, however, I got a totally bad results. This is the reconstruction results after 30k iters:

where the gt mesh is like: I haved adjusted the dataset format and config files, but I don't know whether the result is due to the method or due to some hyperparameters. The adjusted config file is below:

model:
    object:
        sdf:
            mlp:
                inside_out: True
                out_bias: 0.6
            encoding:
                coarse2fine:
                    init_active_level: 8
    appear_embed:
        enabled: True
        dim: 8
    background:
        enabled: False
    render:
        num_samples:
            background: 0

Additinally, I normalize ScanNet scene into the cube of [-0.5, 0.5], and I changed the near/far sampling to 0 and 1.5

Do you have any ideas why ScanNet appears so poor performance?

Dear @zParquet，

I'm curious, did you observe improved results after extending the number of training iterations?

Best, Mulin

[440981]

Is neuralangelo only applicable to 3D reconstruction of a single object from multiple perspectives, rather than 3D reconstruction of internal space? Because I shot a piece of indoor footage, after training, the result is a round object, not an indoor space

[wen-yuan-zhang]

@MulinYu Hi, I'm currently running a new experiment with larger training iterations. I will post the result as soon as the experiment completes ;)

[nanhui69]

chris-aeviator

how to visualize this ply result？

[wen-yuan-zhang]

This time I trained scannet scene0050_00 for 500k iters. It seems better than that of 30k iters, but is still bad :(

I plan to rerun the script using a larger dict size and expect it can become better...

[chenhsuanlin]

@440981 @zParquet have you set the scene type to indoor in the data processing step? Also, you would need to make sure the bounding sphere encapsulates the entire indoor region.

[wen-yuan-zhang]

@chenhsuanlin I transformed the data type manually from raw scannet datset. I rescaled the scene so that the center of the scene lies in the coordinate origin (0,0,0) and the scale of the scene lies in the cube of [-0.5, 0.5]^3. I think this transformation plays the same role as "indoor" processing step. I noticed that neuralangelo initializes the scene as a sphere with radius 1, so I'm convinced that the scene is within the indoor region after transformation.

[chenhsuanlin]

@zParquet we have pushed an update to main yesterday that fixed a checkpoint issue which may be related. Could you pull and try running the pipeline again? Please let me know if there are further issues with the latest code.

[LSK0821]

@zParquet Hello, do you have solved it? I met the same problem, poor performance on Scannet dataset.

[alan355]

This time I trained scannet scene0050_00 for 500k iters. It seems better than that of 30k iters, but is still bad :(

I plan to rerun the script using a larger dict size and expect it can become better...

Have you solved this problem? I'm having the same issue with the indoor dataset as you

[wen-yuan-zhang]

I did not get a satisfied result at last😥. I'm afraid that maybe there is some generalization ability limit of neuralangelo in indoor scene datatset. If anyone can obtain better results, please let me know where I made mistakes.