data_preprocess

Data preprocessing

Our data preprocessing code is largely built upon scripts from Structured3D and scripts provided by the authors of MonteFloor.

Structured3D

• Step 1: download and unzip the Structured3D dataset from here. Note only the Structured3D_panorama folder is required.
• Step 2: convert the registered multi-view RGB-D panoramas to point clouds:

  python generate_point_cloud_stru3d.py --data_root=Structured3D_panorama

• Step 3: project point cloud to density map and generate coco format annotation:

  python generate_coco_stru3d.py --data_root=Structured3D_panorama --output=coco_stru3d

SceneCAD

• Step 1: download 3D scans (<scanId>_vh_clean_2.ply) from ScanNet.
• Step 2: download scan transformation matrics from here.
• Step 3: download 3D layout annotation for ScanNet from SceneCAD.
• Step 4: extract 2D polygon from the 3D layout annotation:

  python extract_2dfloor.py --scannet_planes_path=scannet_planes --scans_transform_path=scans_transform --out_path=2Dfloor_planes

• Step 5: project point cloud to density map and generate coco format annotation:

  python generate_coco_scenecad.py --data_root=ScanNet --scannet_floor_path=2Dfloor_planes --scans_transform_path=scans_transform --output=coco_scenecad

*** Clarify details for SceneCAD ***:

1. Unlike Structured3D, SceneCAD is dominated by single-room scenes. Due to the extreme data imbalance, we filtered out multi-room scenes (which are rare) in SceneCAD. Another motivation is to make evaluation on Floor-SP easier. When evaluating the sequential room-wise shortest path module in Floor-SP, we need to downsample the density map to 64×64 pixels. However, this operation is not proper for multi-room scenes. Please note all methods are benchmarked on the filtered version for a fair comparison.

2. We use slightly different projection methods for Structured3D and SceneCAD. One important difference lies in the quantization operation (see L75-L80 here) in Structured3D. We directly adopted the code from the authors of MonteFloor and only realized this operation in a late stage and didn't change this. In other words, all methods reported in Tab.1 in the paper are trained and tested on the quantized density maps. For SceneCAD, we didn't apply this quantization operation. However, when performing the cross-data generalization experiments (sec. 4.3 in the paper), we evaluate HEAT and RoomFormer on both the original density map and the quantized ones and report the best metrics.