I find ABS_DEPTH_ERROR is close to 6 or even 7 during training, is this normal? Here are the training results for Epoch 5. Is it because of the slow convergence？

avg_test_scalars: {'loss': 4.360309665948113, 'depth_loss': 6.535046514014081, 'entropy_loss': 4.360309665948113, 'abs_depth_error': 6.899323051878795, 'thres2mm_error': 0.16829867261163733, 'thres4mm_error': 0.10954744909229193, 'thres8mm_error': 0.07844322964626443, 'thres14mm_error': 0.06323695212957076, 'thres20mm_error': 0.055751020700780536, 'thres2mm_abserror': 0.597563438798779, 'thres4mm_abserror': 2.7356186663791666, 'thres8mm_abserror': 5.608324628466483, 'thres14mm_abserror': 10.510002394554125, 'thres20mm_abserror': 16.67409769420184, 'thres>20mm_abserror': 78.15814284054947}

Thank you for your great contribution. The script use the gipuma as the fusion method with num_consistent=5，prob_threshold=0.05，disp_threshold=0.25. However, it produces point cloud results with only 1/2 points compared with the point cloud results you provide in DTU, leading to a much poorer result in DTU. Is there any setting wrong in the script? Or because it does not use the dynamic fusion method described in the paper. Could you provide the dynamic fusion process in DTU?

Hi, thank you for sharing this great work!

I'm try to test transmvsnet on tnt advanced dataset, but meet some problem. My test environment is ubuntu16.04 with cuda11.3 and pytorch 1.10.

The first thing is that there is no cams_1 folder under tnt dataset, is it a revised version of original cams folder or you just changed the folder name?

I just changed the folder name, then run scripts/test_tnt.sh, but I find the speed is rather slow, about 10 seconds on 1080ti for a image (1056 x 1920), is it normal?

Finally I get the fused point cloud, but the cloud is meaningless, I checked the depth map and confidence map, all of the data are very strange, apperantly not right.

Can you help me with these problems?

Firstly thanks for your paper and I'm looking forward to your open-sourced code.

And I have some questions about your paper: (Hopefully you can reply, thanks in advance!) (1) In section 4.2, "The model is trained with Adam for 10 epochs with an initial learning rate of 0.001, which decays by a factor of 0.5 respectively after 6, 8, and 12 epochs." I'm confused about the epochs. And I also noticed that this training strategy is different from CasMVSNet. Did you try the training strategy in CasMVSNet? What's the difference? (2) In Table4(b), focal loss(what is the value of \gamma?) suppresses CE loss by 0.06. However, In Table4(e) and Table 6, we infer that the best model use CE loss(FL with \gamma=0). My question is: did you keep Focal loss \gamma unchanged in the Ablation study in Table4? If not, how \gamma changes? Could you elaborate?

Really appreciate it!

Hi, @Lxiangyue Thank you for the nice paper.

It's been over a month since authors announced that the code will be available. May I know when the code will be released? (or whether it will not be released)

Hi thanks for your interesting work. I tested your code on one of the DTU dataset (Moda). as you can see from the following image, the results are quite well.

but I got a very bad result, when i tried to tested on one of my dataset (see the following pic) using your pretrained model (model_dtu). Now, my question is that do you thing that the object is too complicated and different compared to DTU dataset and it is all we can get from the pretrain model without retraining it? is it possible to improve by changing the input parameters? In general, would you please share your opinion about this result?

thanks for your greate work I just tried to do a test on DTU testing dataset I got the depth map for each view but I got a bit confised on how to generate 3D point cloud using your code would you please let me know Best

Hi! Thanks for your excellent work! When I tested on the DTU dataset with pretrained model, the gpu memory consumption is 4439MB, but the paper gives 3778MB.

I do not know where the problem is.

If I have the intrinsic matrics and extrinsic matrics of cameras, which means I don't need to run SFM in COLMAP, how should I struct my data to train the model?

Thanks for the great job. I follow the instruction and upload the reconstruction result of tnt but find the F-score=60.29, and I find the point cloud sizes are a larger than the upload ones. Whether the reconstructed point cloud use the param settting of test_tnt.sh or it should be tuned manually? :smile:

Hi, 我测试了TAT数据集中的Family，使用的是默认脚本test_tnt.sh，采用normal融合，最近仅得到13MB点云文件。经检查发现生成的mask文件夹中的_geo.png都是大部分区域黑色图片,从而最后得到的 final.png的大部分区域都是无效的。geometric consistency阈值分别是默认的0.01和1。不知道您这边是否有一样的问题？

I have tried the pre-trained model you offered on DTU dataset. But the results I got are mean_acc=0.299, mean_comp=0.385, overall=0.342, and the results you presented in the paper are mean_acc=0.321, mean_comp=0.289, overall=0.305.

I do not know where the problem is.