search

hugoycj / Instant-angelo

Instant-angelo: Build high-fidelity Digital Twin within 20 Minutes!
MIT License
426 stars 28 forks source link

Not good performance in an open source dataset #31

Closed city19992 closed 11 months ago

city19992 commented 11 months ago

Nice job. I tested your code on an open source which could be download from https://doc.arcgis.com/en/drone2map/latest/help/sample-data.htm.

But the reconstructed geometry is not so reasonable. Is there any possible reason?

it50000-1 it50000-3

The first one was trained based on neuralangelo-colmap_sparse-50k.yaml. In the second experiment, I centered and scaled the 3D point cloud into NDC space and set the sampling range of ray marching to the outer bounding box of the point cloud. Other settings are the same. But both of the results are not so good.

hugoycj commented 11 months ago

Thank you for the results! I will review the data in more detail

hugoycj commented 11 months ago

Sorry to bother you, I fail to reproduce the issues. The results seem OK on my side.

2 1

hugoycj commented 11 months ago

We have built a docker for this repo, maybe you could run the latest with this environment to see whether it could fix

docker pull hugoycj/instant-angelo
docker run -it hugoycj/instant-angelo
city19992 commented 11 months ago

Thank you so much for the reply. I will try it later.

hugoycj commented 11 months ago

By the way, the current results are generated using the neuralangelo-colmap_sparse.yaml configuration. We have temporarily discarded the neuralangelo-colmap_sparse-50k.yaml since we found limited quality improvement and it is less robust than the neuralangelo-colmap_sparse.yaml. For detailed reconstruction, we sincerely recommend the neuralangelo-colmap_dense.yaml which utilizes the dense point cloud from MVSNet or even from other MVS tools like drone2map as a prior.