Hope to receive your reply! I have some doubts about the point cloud results obtained from testing:
Why are there some black point clouds around the obtained point cloud?

您好，我不知道为什么下载的数据集在测试时会出现这个错误。

Hello, I modified the conf indicator according to what you said, and the final acc, comp, and all indicators are very close to your paper, but the qualitative analysis results are still very different from those in your paper, no matter how small the settings are, there is still a big difference between them and your paper, I haven't changed your code, I don't know what the reason is, can you help solve it, the GPU I use is A5000, or the evaluation of the DTU dataset needs to run test_pcd_dtu.sh instead of test_dtu.sh when I set the conf indicator to 0.1, the effect is as follows.

Hello, thank you very much for your excellent work, I would like to know, for the TNT dataset, how do you get your own pre-trained model, or is it just training on the DTU dataset and testing on the rest of the dataset.

Hello, I have some questions when I reproduce your work, and I hope you can help answer them.

1. After completing the training of the DTU dataset, I tested it in the test set, and the acc, comp, and all indicators were the same as in the article, but there was noise in the obtained point cloud data, so I referred to your reply to others and changed the relevant part of the test_dypcd.py file to photometric_confidence = confidence_list[0]confidence_list[1] confidence_list[2]*confidence_list[3], but after evaluation, it was found that although the noise of the point cloud data is much less, the relevant data indicators are much worse than the original text, can you help with this?
2. For the data in your paper, whether the point cloud data with noise removal is used for evaluation？
   
   

Thanks for this good work! I think you could add a license file to your repo?

Hi, I have two questions.
1、After completing the training process, I have acquired 15 checkpoints ranging from finalmodel_0.ckpt to finalmodel_14.ckpt. I want to select the best checkpoint, however, I've noticed that the evaluation process, particularly the execution time in MATLAB, is quite extensive. Could you provide me with some suggestions?
2、Transformer-based MVS methods, including your work and TransMVSNet, aim to enhance feature representation. However, pre-trained large models can also be used to enhance features. What's your opinion on this?

Thank you for your open-sourcing works!
I noticed that the script uses different depth samples when testing on intermediate (8,8,4,4) and advanced (16,8,4,4) sets of the TnT dataset. Is there any theory or assumption behind using different sets?

Thank you for your response in issue: Modification of point cloud fusion method, however, it does not address my concerns. Is it fair to attribute the DTU score improvement brought about by changing the fusion method to the ET module?

In the baseline results, it seems that you utilized a default fusion method, but for your proposed ET module, there appears to have been a switch to your updated fusion method. This change seems to have led to performance improvements being attributed to the ET module, which may not entirely reflect its true impact. The MVSTER method achieves comparable scores using the updated fusion method alone on DTU, without the need for the ET module. Therefore, attributing the performance gains solely to the Epipolar Transformer in the ablation study seems unfair.

I appreciate your time and attention to these inquiries. I believe that addressing these concerns will not only enhance the credibility of this research but also contribute to the advancement of our understanding of the MVS field.

Traceback (most recent call last):
File "C:\Users\dell\Desktop\ET-MVSNet-main\ET-MVSNet-main\test_dtu_dypcd.py", line 511, in
save_depth(testlist)
File "C:\Users\dell\Desktop\ET-MVSNet-main\ET-MVSNet-main\test_dtu_dypcd.py", line 392, in save_depth
torch.cuda.reset_peak_memory_stats()
File "D:\software\Anaconda\envs\etmvsnet\lib\site-packages\torch\cuda\memory.py", line 260, in reset_peak_memory_stats
return torch._C._cuda_resetPeakMemoryStats(device)
RuntimeError: invalid argument to reset_peak_memory_stats

Hello author, this is the result of my reproduction, it turns out to be my problem, but these final indicators are still a little bit behind your paper, what is the reason for this, is there a way to improve it

Hello, thank you very much for your excellent work, I recently reproduced your work following the tutorial you posted, but there is a little problem, I need your help.

1. The quantitative results in the DTU dataset are basically the same as the results presented in your paper, but the evaluation effect in the TNT dataset is very poor, not even half of the data you present, I would like to ask if there is any good solution. I would appreciate it if you could tell me how
2. Is it possible that there is a problem with my steps, I will now roughly describe it, you see if it is appropriate. Firstly, the DTU dataset is used for training, the trained model is put into the blended dataset for fine-tuning to obtain the blended training model, and finally the blended training model is used as a checkpoint to test the TNT dataset.
   

I followed the instructions on the README and get good outputs for the DTU test set, but when I test on TNT, I just get distorted depth maps, as shown below.

I downloaded the linked TNT dataset and the two trained checkpoints (DTU and finetune on BLD). Both checkpoints fail for me on TNT.

First of all, I want to thank you for your great work and express my gratitude for your open-source code.

The paper mentioned that the baseline is MVSTER. After adding the epipolar transformer, the score on the DTU data set has been greatly improved. However, I found that the dypcd script used in your code is different from that of MVSTER. We know that even for the same depth map, using different fusion methods or fusion parameters, the results will be very different. In your ablation experiments, are you using the exact same fusion script as MVSTER? Would such an experiment be fair if different fusion scripts were used?

I tried deleting the ET module, and as long as I used the "updated" fusion script, I could still achieve similar scores. Here are some of my results. (ET means using fusion scripts from your code repository, and regular means using scripts from MVSTER and other similar methods.)

The changes to the fusion script played a big role in the model's performance on the DTU dataset.

I'm looking forward to your explanation about it, thank you!

Hello, thank you very much for your excellent work, I have recently reproduced your work, but when I was evaluating, I found that there is a lot of noise in the point cloud file, can I reduce the noise by modifying the parameters in the sh file, if so, please inform me which parameter it is, thank you very much.

Thank you for your excellent work. I encountered the following error message while replicating the DTU test dataset. I hope you can answer it and look forward to your reply.

save_depth(testlist)
File "/home/featurize/work/ET-MVSNet-main/test_dtu_dypcd.py", line 392, in save_depth
torch.cuda.reset_peak_memory_stats()
File "/environment/miniconda3/envs/etmvsnet/lib/python3.10/site-packages/torch/cuda/memory.py", line 309, in reset_peak_memory_stats
return torch._C._cuda_resetPeakMemoryStats(device)
RuntimeError: invalid argument to reset_peak_memory_stats

Hello, I reproduced your results on the DTU dataset, but I found that the final training loss is 1.55, I don't know if this result is correct, your paper does not mention the final training loss of the DTU dataset, so it is convenient to inform, in addition, I am ready to reproduce your work, as a baseline, I don't know if it is feasible

Great job! I am interested in how to visualize the epipolar pairs in the diagrams (Fig 1. (a), Fig. 3, and the center of Fig. 4) of your paper. Would you mind sharing the visualization code? I would appreciate it very much.

Hello author, I recently studied your code and came to the following conclusions:

1. Your work mainly focuses on feature enhancement and integrating the ET module into the FPN network.
2. Cost body generation, regularization, and depth estimation are not the focus of your work. Can I make improvements in these aspects in the future?

Hello, I'm sorry to bother you again, it may be that my current level is low, and when I read your paper, I found that the quantitative results on the DTU and TNT datasets you mentioned in the paper, for the test results of the TNT dataset, I can get them by running the two test files you provided? Or do you need to take additional steps, for the data in your paper, do you get it from Tensorboard?

Hi,what an excellent work!After reading your paper,there are still some problems confused me.Would u mind release the code soon or later so that I could have a more comprehensive understanding?Thanks a lot!

Hello, I have two questions to ask

1. For testing TNT datasets, this process is not very time-consuming.
2. Replicate your work and see if there will be any differences in the quantitative results without making any changes

Thanks for the work! I read several works about Epipolar Transofomer, e.g. MVSTER. Homography matrix is able to directly project point p to corresponding p'. Then why apply Epipolar Geometry to find point (p)'s corresponding epipolar line? I've read MVSTER code, they still construct a cost volume from Homography warping.