Intermediate Presentation

Intermediate Presentation

Task Description

1. Get DeepSDF code running and set a baseline.
2. Adjust DeepSDF code to get a new network as mentioned in the DeepLS paper.
3. Train the DeepLS network and get similar results as mentioned in the paper.
4. Evaluate the DeepLS results and compare them with DeepSDF.

What is DeepSDF?

• Network with an AutoDecoder that learns Signed Distance Function representation for a class of shapes.
• Architecture:
  • 8 fully-connected layers with Dropouts and ReLus
  • Last layer tanh

What is a Signed Distance Function?

DeepSDF Results:

Sample Input and Reconstructed Image (From Paper)

Reconstructed Samples (Own Attempt)

How does DeepLS improve DeepSDF?

• Replaces globally learned Signed Distance Function with set of locally learned Signed Distance Functions.
• Architecture:
  • 4 fully-connected layers with leaky ReLUs
  • Last layer tanh

How does the locally learned Signed Distance Function work?

Result comparison DeepSDF vs. DeepLS

Reconstruction (From Paper)

Reconstructed Samples (Own Attempt)

DeepLS

DeepSDF

Problems and next Steps

• Setup of preprocessing step
• Size of Pytorch Embedding
• Correct implementation of Batching