search

autonomousvision / differentiable_volumetric_rendering

This repository contains the code for the CVPR 2020 paper "Differentiable Volumetric Rendering: Learning Implicit 3D Representations without 3D Supervision"
http://www.cvlibs.net/publications/Niemeyer2020CVPR.pdf
MIT License
794 stars 90 forks source link

Scale of point clouds in DTU datasets? #67

Closed jmunkberg closed 2 years ago

jmunkberg commented 2 years ago

Hello, Congrats to a great research project and thanks for releasing the source code!

I'm trying to measure Chamfer loss using DTU scan65, but I'm unsure what scale/coordinate frame the point cloud at https://s3.eu-central-1.amazonaws.com/avg-projects/differentiable_volumetric_rendering/data/DTU.zip DTU\scan65\scan65\pcl.npz is stored in?

I tried computing the AABB around the point cloud DTU\scan65\scan65\pcl.npz:

-1046.1594 197.94109
-898.51086 832.34924
146.50441 904.0597

Now. when running python generate.py configs/multi_view_reconstruction/skull/ours_rgb_pretrained.yaml I get two meshes out: scan65.ply with AABB within [-0.5, 0.5] and scan65_world_scale.ply with AABB

-161.42103576660156 68.33500671386719
-154.46188354492188 93.75101470947266
542.672607421875 749.3994140625

I assume it's the latter version that should be used in the Chamfer computation, but there is a large difference in the AABBs. Am I missing some transform, or are there other reference point clouds I should use?

Also, what normalization factors (maximal edge lengths) should I use for the three DTU examples, following (Like Fan et al. [17] we use 1/10 times the maximal edge length of the current object’s bounding box as unit 1.)?

jmunkberg commented 2 years ago

NVM, I missed the `'is_in_visual_hull field.