search

TQTQliu / ET-MVSNet

[ICCV 2023] When Epipolar Constraint Meets Non-local Operators in Multi-View Stereo
MIT License
57 stars 2 forks source link

Trained checkpoints and feature enhancement approaches #10

Open Warrior456 opened 4 months ago

Warrior456 commented 4 months ago

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?

TQTQliu commented 4 months ago
  1. You can select the checkpoint based on the depth error, like abs_depth_error at each epoch. But it's worth noting that a smaller depth error does not necessarily guarantee better 3D reconstruction metrics. According to my experimental experience, you could choose one from finalmodel_12.ckpt to finalmodel_14.ckpt. For evaluating reconstruction metrics, apart from using standard MATLAB, you can also employ a Python implementation as a reference. However, it is important to note that there are some gaps in the performance results obtained by these two methods.
  2. Leveraging pre-trained large models can effectively boost a model's feature representation capability and generalization. I recommend that you refer to MVSFormer and MVSFormer++ for further insights in this regard.
05063112lcs commented 1 week ago

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? Hello, I saw the question you posted in the ET-MVSNet project, I have a question that I need to ask, can you reproduce the quantitative results of the TNT dataset in the paper, why the indicators I got are very poor, very poor.