Download the BlendedMVS dataset via this link.
For now, we have tested on the low-res set
of the BlendedMVS
split (27.5GB).
This is only for the original BMVS dataset. We have provided a small tested split with already normalized cams. See links above.
The camera poses in BlendedMVS dataset have random centers and random scales. Therefore, it is necessary to first normalize the camera centers to the origin and standardize the scale of the camera distances.
In order to achieve this, we have developed a simple script to find the focus center of multiple object-centric views, by minimizing the average projected pixel distance w.r.t. the image centers across all frames. This script only requires camera intrinsics and poses.
cd /path/to/neuralsim
python dataio/bmvs/normalize_bmvs.py --root /path/to/bmvs
It will generate cameras_sphere.npz
files in the instance directories:
/path/to/bmvs
├── 5aa0f9d7a9efce63548c69a1
│ ├── cameras_sphere.npz
│ ├── blended_image
│ │ ├── 00000000.jpg
│ │ ├── 00000001.jpg
│ │ ├── ...
│ ├── cams
│ │ ├── 00000000_cam.txt
│ │ ├── 00000001_cam.txt
│ │ ├── ...
├── 5aa235f64a17b335eeaf9609
│ ├── cameras_sphere.npz
│ ├── ...
├── ...
cd /path/to/neuralsim
source set_env.sh
python dataio/bmvs/bmvs_dataset.py