计算机视觉

• Image(video) -> (sensing device) -> (interpreting device) -> Interpretations
• Automatic understanding of images and video
  • Measurement: Computing properties of the 3D world from visual data
  • Perception and interpretation: Algorithms and representations to allow a machine to recognize objects, scene and people
• 应用
  • Faces and digital cameras
  • Video-based interfaces
  • safety and security
    • Navigation, driver safety
    • Monitoring pool
    • Pedestrian detection
    • Surveillance
  • Vison for medical and neuroimages
• 困难：Gap between low level signal and high level meanings
  • ill posed problem
  • large variation: illumination, object pose, clutter, occlusions, intra-class appearance, viewpoint
  • intra-class variation
  • context
  • complexity
• Progress chart by dataset
  • Roberts 1963
  • COIL
  • MIT-CMU Faces (2000)
  • UIUC Cars
  • INRIA Pedestrians
  • MSRC 21 Objects (2005)
  • Caltech-101
  • Caltech-256
  • PASCAL VOC (2010)：奠定了计算机视觉评价体系
  • Faces in the Wild
  • 80M Tiny Images
  • ImageNet：多；长尾
  • Birds-200
• Tasks
  • Recogintion: General categories
  • Large scale recognition
  • Recognition in first-person view
  • Object detection, instance segmentation
  • image captioning
  • image generation
• CVPR'19 by the numbers
  • Submissions: 5160
  • Accepted: 1294
  • Registered Attendees: 9227
• Marr’s vision framework
  • computational level
  • algorithmic/representational level
  • implementational/physical level
• Malik’s Perspective: Recognition, Reorganization, Reconstruction
• Important note: In general, computer vision does not work (except in certain situations/conditions)