search

JiayuYANG / CVP-MVSNet

Cost Volume Pyramid Based Depth Inference for Multi-View Stereo (CVPR 2020 Oral)
238 stars 33 forks source link

How to choose the specific model from 40 models #13

Closed etudemin closed 4 years ago

etudemin commented 4 years ago

Hi @JiayuYANG,

Thanks for your great work!

In this project, authors provided the pre-trained model of the 27th epoch (model_0027.ckpt). Would you kindly explain how the authors choose the specific model from 40 models?

I've retrained your network using the default setting, and got 40 models which correspond to 40 epochs. Attached is the generated log file, and the epoch loss is dropping gradually during 40 epochs. Seems that there is no validation process during training, so it's a little confusing about how to choose the specific model to test.

Thanks for your kind reply in advance, and please correct me if I miss something. 20200831-0132.log

JiayuYANG commented 4 years ago

Hi @etudemin ,

You can modify my eval.sh to use the validation set provided by MVSNet to calculate depth errors if you would like to do validation on depth maps.

And after that you can also use fusion.sh to generate validation point clouds and compute the accuracy and completeness scores using DTU evaluation code if you would like to do validation on point clouds.

Cheers, Jiayu

etudemin commented 4 years ago

Hi @JiayuYANG ,

I will go check the validation set. Thank you very much for your suggestions!

Sincerely, etudemin