tedyhabtegebrial / SoftOcclusionMSI

Code accompanying CVPR 2022 paper, " SOMSI: Spherical Novel View Synthesis with Soft Occlusion Multi-Sphere Images"

SOMSI: Spherical Novel View Synthesis with Soft Occlusion Multi-Sphere Images

Tewodros Habtegebrial , Christiano Gava , Marcel Rogge , Didier Stricker, Varun Jampani
IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR) 2022
The project page can be found here

Dependencies

This code was develped with python3.6

pytorch_lightning==1.2.4
torch==1.7.1
numpy==1.18.5
opencv-python==4.5.1
kornia==0.4.1

Datasets

Download link. Save the dataset to the folder somsi_data Please check the datasets page for more details on each dataset.

Training your own models

Please check the following script as an example.

./scripts/med_port.sh

Testing our pre-trained models

Download models from this link. Extract the downloads to ```ckpts`` directory under the project directory.

Take a look at a sample test script. Before running, check the note on ERP resolution and appearance feature size, below.

./scripts/med_port_test.sh

The following parameters are crucial to avoid errors.

ERP Resolution

We have trained models with MatryODSHka resolutoin $320 \times 640$ and $768 \times 1536$. Make sure the ckpt you are using and the height and width parameters match. Ckpt folders end with $height \times width$.

Appearance Feature Dimensionality

We provide models with the following features sizes 3, 12, 24

If the model ckpt base file name contains feat_x $x \in [3, 12, 24]$, pass

--feats_per_layer=x # x in [3,12,24]

Scene Number

Replica and Residential Area datasets have, 14 and 3 scenes, respectively. During training and testing pass the corect scene_number parameter. scene_number ranges from 0-13`` for Replica and 0-2``` for Residential Area datasets.

Missing Features

• Test Code
• Demo Code
• Datasets
  • Residential
  • Replica
  • Coffee Area
• Reflectance Modelling
• S-NeRF Baseline

Acknowledgments: This repo builds upon the Nerf-PL.