3D reconstruction and modeling of urban buildings and architecture

Objective

  • To realize automatic/semi-automatic reconstruction and modeling of urban building or architecture with complex components

Challenges

  • Photo-realistic & geometrically accurate
  • Multi-images of target from different vantage points may not always be available

Approach

Employ constraints derived from shape symmetry

  • Camera calibration by exploiting symmetry
  • Duplication of virtual cameras
  • Reconstruction by combining real and virtual pairs
  • Modeling according to shape symmetry

Vision-based human activity analysis

  • Efficient analysis of human actions and activities in different scenarios based on visual data.

Trajectory-based motion representation for action recognition

Objective

  • To recognize common (single or dual actor) actions in natural video sequences

Challenges

  • To extract motion information in video
  • To model temporal motions and their relationships

Approach

  • Motion information encoded in motion trajectories
  • Hierarchical levels of info extracted: appearance, intra-trajectories, and inter-trajectories

Contextualized histograms for group activity analysis

Objective

  • To analyze group activities that involve multiple actors with interactions

Challenges

  • To model interaction between different actors
  • To keep the model complexity low

Approaches

  • Raw motion information encoded in higher order local statistics
  • Randomized forest employed to facilitate model simplification with supervision