To generate bird’s-eye-view (BEV) images and project 3D object marks, you need to download three indoor scene datasets:
After downloading, organize the datasets in the following directory structure:
.
├── arkitscenes
│ ├── data
│ │ ├── ...
│ │ └── {scene_id}_3dod_mesh_wo_ceiling.ply # point cloud files
│ ├── gt_annotation # ground truth annotations
│ │ └── {scene_id}_3dod_annotation.json
│ └── detection_annotation # noisy detection annotations
│
├── scannet
│ ├── data
│ │ ├── ...
│ │ └── {scene_id}_vh_clean_2.ply
│ ├── gt_annotation
│ │ ├── {scene_id}.aggregation.json
│ │ └── {scene_id}_vh_clean_2.0.010000.segs.json
│ └── detection_annotation
│
├── scannetpp
│ ├── data
│ │ ├── ...
│ │ └── {scene_id}_mesh_aligned_0.05.ply
│ ├── gt_annotation
│ │ ├── {scene_id}_segments_anno.json
│ │ └── {scene_id}_segments.json
│ └── detection_annotation
└── ...
Use the script below to generate BEV images and save 3D object bounding boxes as JSON files:
python generate_bev_images.py --dataset all
We are currently streamlining the data generation pipeline for training and evaluation, as the original datasets are too large to be released directly.