This is the official implementation for Multi-task Planar Reconstruction with Feature Warping Guidance, published at VISAPP 2024.
Base Model | Feature warping |
---|---|
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
│ └── ...
└── ...
conda create -n testsolop python=3.10
pip install -r requirements.txt
Download resnet-50 pretrained backbone here and put in pretrained/
folder
Edit config files with correct dataroot and settings
python train_local.py
Download our pretrained model here
Update the corresponding config file model.path
python eval_models.py
- Modules adjusted from: https://github.com/OpenFirework/pytorch_solov2
- ScanNet dataset annotations from: https://github.com/NVlabs/planercnn