《Shape Completion using 3D-Encoder-Predictor CNNs and Shape Synthesis Angela》

pipeline

graph TD
A[3D-EPN: predict global structure in unknown area] --> B[correlate these intermediary result with 3D geometry from a shape database]
B --> C[patch-based 3D shape synthesis method]

Contribution

• 3D-EPN completes partially-scanned 3D models while using semantic context from a shape classification network.
• 3D mesh synthesis procedure to obtain high-resolution output and local geometry details
• end to end completion method

shape completion

1、clean up for broken 3D models
laplacion smooth and poission reconstrcutin is in
the category

2、detecting structure and regularities in 3D shape
limits the shape space in hand-crafted design

3、Much research leverages strong data-base priors
limits cannot generalize to new shapes
there key insight is take these information for global structure rather than local information

4、fully data-driven method trained with machine learning techniques is promising direction

Voxlets

They train a random decision forests that predict unknown voxel neighborhoods; the final mesh is generated with a weighted average of the predicted results

3D ShapeNet

They also use convolutional neural networks – specifically a deep belief network – to obtain a generative model for a given shape database.
this strategy is significantly less efficient than directly training an end-to-end predictor as our 3D-EPN does