Poor results (mesh and rendered normals) on Courthouse

[shayan-kousha]

Hi, Thanks for sharing the wonderful work!

I'm trying to reproduce the results shown on the project page on Courthouse. However, the mesh and rendered normals that I get after training the model for 700k iterations are significantly worse than the results on the project page.

I've trained the neuralangelo model on Courthouse using the provided tnt.yaml. The only modification I made to the config file is changing num_images to 1106 as Courthouse has more training images than Barn. The data is processed by following the steps in DATA_PROCESSING.md.

Here are some results after training the model for 700K iterations: Rendered RGB image:

Rendered normals:

Mesh:

I'd appreciate it if you could help me improve these results and reproduce the mesh and normal rendering shown on the project page.

[mli0603]

Hi @shayan-kousha

I am looking into potential bugs in the pipeline that may have also caused several other issues reported. I will update once confirmed.