SOLOPlanes

This is the official implementation for Multi-task Planar Reconstruction with Feature Warping Guidance, published at VISAPP 2024.

Base Model	Feature warping

Dataset

Download the ScanNet dataset and follow instructions to extract sens data for RGB, depth, and camera parameters.

We use plane annotations from https://github.com/NVlabs/planercnn#training

The data directory should look like:

ScanNet
├── scans
│   └── sceneXXXX_XX
│       ├── annotation 
│       ├── sens    
│       └── ...
└── ...

Getting started

conda create -n testsolop python=3.10
pip install -r requirements.txt

Training

Download resnet-50 pretrained backbone here and put in pretrained/ folder

Edit config files with correct dataroot and settings

python train_local.py

Evaluation

Download our pretrained model here

Update the corresponding config file model.path

python eval_models.py

Acknowledgement

• Modules adjusted from: https://github.com/OpenFirework/pytorch_solov2
• ScanNet dataset annotations from: https://github.com/NVlabs/planercnn