Introduction

Links:

Visual Domain Adaptation Challenge:

  • Segmentation Winner: MSRA
  • Segmentation honorable mention: NEC
  1. Feature Level
    • align the features extracted from the networks across the sourse and the target domains. (Unsupervised: no labeled target samples)
    • typically, minimize some measure of the distance between the source and the target feature distribution,
      • maximum mean discrepency:
      • Correlation distance
      • adversarial discriminator accuracy
    • Limitaions:
      1. align marginal distributions does not enforce any semantic consistency. (e.g. car->bicycle) If the feature distributions are quite different??
      2. higher levels of a deep representation can fail to model aspects of low-level appearance variance lose some low-level/local feature/information
  2. Pixel/Frame Level : Generative
    • similar distribution alignment. Translate the source data to the style of a target domain. similar distribution alignment: If the feature distributions are quite different??
    • Unsupervised methods:
    • Limiation:
      • small image sizes and limited domain shifts;
      • controlled enveironment;
      • may not preserve content: crucial semantic information may be lost
  3. Multilevel
  • Discriminator:

    • cannot distinguish the segmentation result between source and target
  • How to use the images only with the view point changes

  • The distributiion of the features are very different from each other
    • Also, scale is different.
  • The categories of source and target are not exactly same.

results matching ""

    No results matching ""