yiqun-wang / hfs Goto Github PK

Hi,
Before all, thank you for your inspiring work !
I can get the mesh file through this job, I would like to know what tool you use to get the water-tight mesh.
for example：

Hello authors!

When reading through section 3.1 of the paper, it's a bit unclear how to handle alpha composition and what parts exactly are taken from NeuS to handle the multiple surface intersection case.

Could you please elaborate on those missing details? (i.e how exactly everything works in the discretized setting)

Thank you!

Dear author, this is a great job, but I encountered a problem. When I wanted to change from single card training to multi card training, my GPUs did not train together. Do you know the reason or can you provide a solution? Thank you very much for your answer

your work is really nice, and I have a question about dataset, do you have blender dataset? such as "chair" ..., your google driver only have images, but I want to eval the mesh, I need bounding box and mesh

The names and formulas of these two variables confused me.
s = deviation_network(torch.zeros([1])).clip(1e-6, 1e3).detach()
inv_s = deviation_network(torch.zeros([1, 3]))[:, :1].clip(1e-6, 1e6)

Congratulations on your amazing results and acceptance for Neurips 2022, I wonder what is your code releasing plan, eg. when will you release the code? Can't wait to use your amazing projects! Thank you.

Hi,

I tried to run HFS on some custom datasets generated using blender (same style as the synthetic dataset used by NeRF). Most of them worked properly, with only a scene appears to be very weird -- it has okay-ish coarse and fine rendering, but the normal rendering and mesh extracted are all just empty (see below)

Does it mean that the model failed to reconstruct the proper surface (probably because in this scene the objects are disconnected?) and models everything using the background model?

Hi, yiqun!

I'm interested in using the DTU training config with mask, but I noticed that it has not been released yet. Would it be possible to provide the config file?

Thank you for your open source contributions and I appreciate your help.

Best regards

Hi,

Congrats on your great work! I am currently running on BlendedMVS and cannot find the evaluation code for it. If convenient, could you provide your eval code for BlendedMVS?

Thanks!

Great job! Can you share the code used to evaluate mesh for NeRF synthetic datasets? In issues 3, I see five shared obj files. Do you have gt mesh for the other three scenes?

Hi authors,

First of all, thanks for your great work!
I am testing the mesh reconstruction quality with/without the displacement function(or base sdf versus combined sdf). I set weaken to 0 in fields_high.py to generate the mesh without the displacement function. However the meshes look almost the same. It seems that no extra details are added, only the surface is slightly moved along its normal.

I noticed that in your supplementary material the difference of extra details seems to be not negligible. Have I made something wrong in mesh reconstruction, or it's exactly the desired results?

Has anyone encountered this problem?

Hello Wooden
Load data: Begin
Traceback (most recent call last):
File "exp_runner_high.py", line 503, in
runner = Runner(args.conf, args.mode, args.case, args.is_continue, args.ckpt_name, args.base_exp_dir, args.end_iter)
File "exp_runner_high.py", line 51, in init
self.dataset = Dataset(self.conf['dataset'])
File "/root/autodl-tmp/HFS/models/dataset.py", line 79, in init
self.intrinsics_all_inv = torch.inverse(self.intrinsics_all) # [n_images, 4, 4]
RuntimeError: CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasCreate(handle)

Thanks for open-sourcing your work with the greater community. The recovered high-frequency details are indeed noticeable, albeit with extra training latency. One other issue I'd like to highlight, and hopefully get your input on is the GPU memory overhead which seems to scale rather intensely with image resolution and number.

Is there a way to mitigate this memory overhead for larger datasets without trading off too much of accuracy? The rays/batch size does not seem to help in this regard.

Thanks for the great project! Can you provide training data for NeRF_synthetic scenes and DTU scenes?

Hi,

Before all, thank you for your inspiring work !
I'm trying to process some custom data (that I preprocessed with COLMAP as indicated in the original NeuS repository) and compare the result with plain NeuS but I feel like the training process takes much more time than it should be (actually ~5*training time of plain NeuS).
I have strange outputs from TQDM in the logs. It seems that it cannot estimate the true duration of an iteration. I get 7 digits it/s which is obviously wrong and consequently the remaining time metric is wrong as well (0 s. remaining).
Obviously training an additional MLP for the second SDF should make it slower to train but I didn't expect such drawback...

I was wondering if it was 'normal' as in you're aware of the situation or should I consider this as unusual and investigate my data ?

Thank you in advance!

Hello,
Thank you for this great work. Could you please release the code that uses the pre trained model? Thank you very much.