Official dataset for Semantic Implicit Neural Scene Representations With Semi-Supervised Training: https://arxiv.org/abs/2003.12673
Code coming soon!! For now please refer to https://github.com/vsitzmann/scene-representation-networks
Please run setup.sh in order to download the dataset. It will create semantic_srn_data/ and download the Chair/ and Table/ directories from box.com into it.
The main directory contains all object directories (e.g., Chair)
.
├── semantic_srn_data
├── Chair
├── Chair.train
├── Chair.val
├── Chair.test
├── Chair.txt # Shows hierarchy of parts (not used)
└── Chair-level-1.txt # Maps class labels (e.g., 0,1,2,..) to part names (e.g., chair back, chair arm, etc).
└── Table
├── ... # Same as Chair
Each split (e.g., Chair.val) contains a number of object instances (e.g., 800dd8ed32104151a37f3fc191551700)
├── ...
├── Chair.val
├── ...
├── 800dd8ed32104151a37f3fc191551700 # Single object instance (e.g., lawn chair)
└── ...
└── ...
Within each instance are the following elements:
.
├── ...
├── 800dd8ed32104151a37f3fc191551700 # Single object instance (e.g., lawn chair)
├── intrinsics.txt # Camera intrinsic parameters used to render the images in rgb and seg
├── point_cloud # Contains point cloud data of the instance (not used)
├── sample-points-all-pts-nor-rgba-10000.txt # point cloud stored as (x y z Nx Ny Nz R G B A)
├── sample-points-label-10000.npy # segmentation class labels for the above point clouds.
├── result_after_merging.json # maps Partnet mesh indices to part names (not used)
├── result.json # maps Partnet mesh indices to part names for point clouds (not used)
├── pose # camera extrinsic parameters used to render each image in rgb and seg.
├── ...
├── rgb # rgb images at each camera view, saved as .png files
├── ...
└── seg # segmentation maps at each camera view, saved as image-shaped numpy (.npy) arrays with per-pixel class labels
├── ...
└── ...
Please direct any questions to apkohli@berkeley.edu