This is the 3D Object Recognition with Deep Networks Project for the 3D Vision course at ETHZ

#####Input Data/Voxel/Occupancy Grid

• read 3D CAD Voxel Data transform to Occupancy Grid
  • ModelNet10 - Zip Datei
    • in VoxNet paper an adapted version is used
  • ModelNet40 - Zip Datei
  • .OFF Files can be view in MeshLab
  • Matlab Function to read .OFF files (Matlab File)
  • Function in the 3D ShapeNet Source code to transform to occupancy grid is called ...
• optional 2.5D Reconstruction (combine multiple 2.5D Representation into new 3D Representation)
• optional 2.5D & 3D Point Cloud Data to Voxel Data (Project Tango & Extra Training Data Sets)

#####VoxNet

• VoxNet Source Code Python - Github
• Convolutional Neural Network
• Input Data
  • Rotation Augementation & Voting - Increases Performance when applied
  • Multiresolution Input (not used on ModelNet)
  • 3 Grids - Best Results for Density Grid
    • Density Grid
    • Binary Grid
    • Hit Grid
• Training: Stoastic Gradient Decent with Momentum
  • learning rate = 0.0001
  • momentum parameter = 0.9
  • batch size = 32
  • learning rate decrease: 10 per 40000 batches.
• Dropout Regularization after output of each layer
• Initialization:
  • Convolutional Layers:
    • forward Propagation: zero mean Gaussian with std.dev = sqrt(2/n_l), n_l = dimension(input array(30x30x30) layer l) * input channels layer l
    • backward Propagation: zero mean Gaussion with std.dev = sqrt(2/n*_l), n*_l = dimension(input array(30x30x30)layer l) * input channels layer l-1
  • Dense Layers: zero-mean Gaussion with std.dev=0.01
• augement training data through randomly pertubed(mirrored/shited) copies
  • mirror along x and y axis
  • shift by -2 to 2 voxels along x and y axis

#####3D Shape Net

• 3D ShapeNet - Source Code Matlab - Zip
  • Special Learning Algorithm
  • Convolutional Deep Beliefe Network

1. Get ModelNet 3D CAD Data & Read & Transform if necessary
2. Build VoxNet
3. Build ShapeNet
4. Train, Validate & Tune

• Not Sure if Cross Validation works with the mix of 3D Training and 2.5D Validation Data, but i guess there is someway
• Depending on how much parameters are left to tune and how much GPU power we have this might take a while (it took them 2 days for 40 objects on a cluster)

5. Test on Spare Data and try to recreate Experiment from Wu et al. & Maturana et al.
6. optional Maybe try to get some Point Cloud Data from Tango Tablet and test Models against that

• Documentation
• Point Cloud to SuperVoxel
• PointCloud Python Support(bad)

• API Documentation
• Tutorials
• TensorFlow GitHub