Hello. Thank you for your amazing work and for sharing the code! I tried training your model from scratch on some of the benchmark scenes and had problems reproducing your results. It seems like the model is quite susceptible to the random seed, and even after several attempts, the quality I obtained was lower than reported.

| Experiment | Fountain - Full | Fountain - Center | Herzjesu - Full | Herzjesu - Center | | :--- | :----: | :----: | :----: | :----: | | pre-trained | 7.77 | 1.91 | 8.88 | 2.03 | | seed 2022 | 8.03 | 2.65 | 7.66 | 2.55 | | seed 42 | 13.36 | 7.43 | 10.54 | 2.58 |

Could you provide some insights regarding this issue?

Follow the instruction of this repo. I got a result, 0.72, on DTU scan24 scene, which is different from the result in this paper, 0.49. Could you please tell me if anywhere I made some mistakes?

Hello,

thanks for your excellent work! I am wondering how you create the result visulization teaser.gif in readme, with a moving camera and a sliding bar to switch between two meshes. Thank you in advance!

Best regards

Dear @fdarmon

Thank you very much for your work. In your error visualization in appendix, the GT points look very clean compared to the provided GT. I wonder if you preprocess the GT points from DTU to get such visualization.

Look forward to your reply!

As descripted in your paper, "the results for each method are taken from their original paper". But why results for Neus[34] are different from original paper?

[34] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Adv. Neural Inform. Process. Syst., 2021.

Dear author, Thank you for your open source code. I notice that volsdf normalize the camera (within a sphere with R=3) on DTU dataset. However, NeuralWarp loads camera params from the original cameras.npz, which is not normalized. Since this project is built upon volsdf, I want to know why the author removed the camera normalization step. Moreover, I think the default value of --bbox_size in extract_mesh.py is not much reasonable. It should be set carefully for each scene because the camera is not normalized. If the value is too large, then a lot of empty space is wasted. If the value is too small, the extracted mesh is defective. I want to hear the author's viewpoint on this. Thanks!

Hello, I encountered OOM Error after "generate_visu" on RTX3080, can you provide some suggestions on how to reduce batch size or use multiple GPU?

Dear author, According to the description of the paper, the training pipeline includes two stages. First, train for 100k iterations in the same setting as VolSDF. Then finetune for 50k iterations with the proposed method. Does this mean that I need to add the option "--is_continue --timestamp XXXXX" in stage 2? Moreover, according to the paper, the learning rate of stage 2 is 1e-5, which is different from the learning rate (5.0e-4) in NeuralWarp.conf. Do I need to change the learning rate in the configuration to 1e-5? Thanks!

Hi, Thanks for the excellent paper.

I was going through the code base and the warping seems to happen only for random pixels sampled on the image plane (viz. clipped off at boundaries), and not patches. line 155-166 in https://github.com/fdarmon/NeuralWarp/blob/main/datasets/scene_dataset.py

Is it right, or am I missing something ?

Thank you very much for the author's contribution, I want to try to train my own data, but I am confused about the pair.txt file in "dtu_supp", how do I get this?

Hello!

Thanks for sharing the code, your work is really impressive! I have noticed that you did not include any license file. Per copyright laws, we have to assume the most restrictive license (i.e. all derivatives are forbidden, etc.)

Was that intentional, or do you actually not mind other research projects building on top of your code?

Hi, your great work is impressing! But I'm a little confused about the warp loss and Validity masks: 1."our warping-based loss such that every valid patch in the reference image is given the same weight." why is same weight? 2."second, when the reference and source views are on two different sides of the plane defined by xi and the normal ni; third, when a camera center is too close to the plane defined by xi and the normal ni." Why the binary indicator V will be 0. Looking forward to your reply.