Data Formats

For 32³, we have train and test data as hdf5. We also provide the raw scanned and complete data at 32³ and 128³ in custom binary formats:

• Partial Data (*.sdf):
  
    #binary
    dimX    #uint64
    dimY    #uint64
    dimZ    #uint64
    data    #(dimX*dimY*dimZ) floats for sdf values


• Complete Data (*.df):
  
    #binary
    dimX    #uint64
    dimY    #uint64
    dimZ    #uint64
    data    #(dimX*dimY*dimZ) floats for df values

Input Data

We provide virtually scanned partial models as from ShapeNet Core, as well as their corresponding distance transforms of the complete models. Files are structured [class id]/[model id]__[trajectory id]__.[sdf/df]

• h5_shapenet_dim32_sdf.zip (30GB)
• shapenet_dim32_df.zip (4GB)
• shapenet_dim32_sdf.zip (11GB)
• shapenet_dim64_df.zip (30GB)
• shapenet_dim128_df.zip (256GB)

Point cloud representations of the partial scans:

• shapenet_dim32_sdf_pc.zip (10GB)

The trajectories for each shapenet model can be found here: shapenet_cameras.zip (115MB). The trajectories for shapenet images test set can be found here: shapenet_cameras_images.zip (321KB).

Benchmark

We provide two synthetic test benchmarks of 1200 partial models each (shapenet model id list here). The images benchmark contains models scanned with a single depth image using a horizontal camera, while the scans benchmark contains models scanned from a trajectory with at least one depth image. We also provide a real test benchmark on real scan data based on the dataset from Qi et. al. 2016, containing instances from the chair, desk, nightstand, sofa, and table categories.

• h5_test-real_dim32_sdf.zip (5MB)
• test-real_dim32_sdf.zip (4MB)
• h5_test-images_dim32_sdf.zip (150MB)
• test-images_dim32_sdf.zip (18MB)
• h5_test-scans_dim32_sdf.zip (180MB, see input data above for corresponding
• test-images_dim128_sdf.zip (960MB)

Point cloud representations:

• test-images_dim32_sdf_pc.zip (13MB)
• for test-scans, see point cloud data above for the respective shapenet model ids

Our results:

• output-test-real-32.zip (10MB)
• output-test-real-128.zip (100MB)
• output-test-images-32.zip (110MB)
• output-test-images-128.zip (800MB)
• output-test-scans-32.zip (110MB)
• output-test-scans-128.zip (800MB)

Trained models (including classifier): trained_models.zip (740MB)

	ℓ1 norm to ground truth distance field (masked)
method	scans, 323	scans, 1283	images, 323	images, 1283
epn-unet-class + synth [1]	0.309	1.80	0.374	1.89
epn-unet + synth [1]	0.310	1.82	0.379	1.91
epn-class + synth [1]	0.376	1.92	0.483	2.16
epn + synth [1]	0.382	1.94	0.512	2.33
3d ShapeNets [2]	-	-	0.905	3.70
ShapeRecon [3]	-	-	0.970	4.63
Poisson Surface Reconstruction [4,5]	-	-	1.91	8.46

To add your results to the leaderboard, please email Angela Dai.

[1] A. Dai, C. Qi, M. Nießner. Shape Completion using 3D-Encoder-Predictor CNNs and Shape Synthesis. CVPR 2017.
[2] Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, J. Xiao. 3D ShapeNets: A Deep Representation for Volumetric Shapes. CVPR 2015.
[3] J. Rock, T. Gupta, J. Thorsen, J. Gwak, D. Shin, D. Hoiem. Completing 3D Object Shape from One Depth Image. CVPR 2015.
[4] M. Kazhdan, M. Bolitho, H. Hoppe. Poisson Surface Reconstruction. Eurographics Symposium on Geometry Processing 2016.
[5] M. Kazhdan, H. Hoppe. Screened Poisson Surface Reconstruction. SIGGRAPH 2013.