Weekly Report 15 (07.22. ~ 07.26.)

Overview

• Timeplan for IntersectNet
• Regular Dataset
• CG Dataset

Timeplan for IntersectNet

• Generate & Prepare Dataset ~2-3 Weeks (Week 16-18) (07.29.~08.16.)
• Train & Fit Network ~2 Weeks (Week 19-20) (08.19.~08.30.)
• Write Report ~1 Week (Week 21) (09.02.~09.06)

Regular Dataset

• 2D-3D-S Dataset
  • 1.413 equirectangular Images
• 360-Dataset
  • superset/derivative of multiple datasets (including 2D-3D-S)
  • 12.072 scanned and 10.024 CG scenes
• Advantages:
  • big, independent dataset
• Disadvantages:
  • amount of corridor scenes/images unknown
  • sorting for dataset might take too long
  • camera position too high in most images

CG Dataset

• Python Script to generate random corridors
• Render euqirectangular surround images at random positions in the corridor
• Advantages:
  • needed amount of images can be generated
• Disadvantages:
  • quality of dataset heavily dependent on quality of textures
  • variation in images may be insufficient and is based on variation of textures

CG Image Generation

• fixed corridors with random centerpiece
• position camera at semi-random location inside corridor and render equirectangular image