• Ehsan Qasemi, qasemi@wisc.edu
• Erfan Khoram, ekhoram@wisc.edu

University of Wisconsin-Madison

Problem Formulation

• Estimation of the spatial structure of a scene
• Essential component in the field of:
  • Understanding the 3D geometry of the scene
  • Unmanned vehicle and robotics
  • Human body pose estimation

Previous Works

• Supervised Learning with Ground Truth Depth dataset
  
  • Use semantic segmentation (Ladicky et al.)
  • Over-segmented patching (Saxena et al.)
  • CNN to learn unary and pairwise terms (liu et al.)
  • Global predictions refined with local information (Eigen et al.)


• Unsupervised Depth Estimation
  • Deep stereo with multiple images (Flynn et al.)
  • Deep 3D: generate right image from left (Xie et al.)
  • LR consistency (Godard et al.)

Future Works

• Use different scales in discriminator
• Comparison Problem in GAN (Vondrick et al.)
• Possible use in 2D to 3D conversion
• Utilize temporal information in video

Any Questions?