Full dataset.
Download the complete SynthCity dataset as a single parquet file. This is not split into seperate areas (27.5 GB).
Areas 1-9.
Download the entire SynthCity dataset split in 9 geographical areas. Each area will fit into memory of a standard work station with RAM of 32GB. The download consists of a .zip containing 9 .parquet files.
Area 1 sample.
Download a small sample (~1/9) of the full dataset in .parquet format (XGB).
Test area only (area 3).
Download the proposed test area of the dataset (area 9) in .parquet format (XGB).
Lidar trajectory file.
Download the motion file for the synthetic lidar at each key frame. The file has contains x,y,z,yaw,pitch,roll as well as time so that it can be easily registered to each individual point or scan line.
PointNet++ trained weights.
As one of the leading end-to-end point cloud architectures we provide the trained weights for PointNet++ as implemented here (tensorflow).
To generate the training data we split the data into 10x10 metre chunks and then applied incremental voxel downsampling until we reached 8192 points. Other hyperparametes include:
Batch size: 16
Learning rate: 0.001
Momentum: 0.9
Decay rate: 0.7
If you would like to obtain the preprocessed data (.npy) please contact us.