Permalink
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
36 lines (29 sloc) 2.49 KB

Learning to Localize Sound Source in Visual Scenes

Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, In So Kweon, CVPR, 2018

Keywords:

sound source localization; cross-modal supervised network; attention model

Summary

  1. Contributions:
  • Introduce a learning framework to localize sound source using the attention mechanism, which is guided by sound information, with a paired sound and video frame;
  • Propose a unified end-to-end deep convolutional neural network architecture that accommodates unsupervised, semi-supervised, and fully-supervised learning;
  • Collect and annotate a new sound source localization dataset, which provides supervised information and facilitates quantitative and qualitative analysis.
  1. Proposed Approach:
  • Design a neural network to address the problem of vision based sound localization within the unsupervised learning framework;
  • The network composes of three main modules: sound network, visual network and attention model.
    • Sound Network: The input is a 1-D signal with varying temporal length, the output is 512-D.
    • Visual Network: Composed of the image feature extractor and localization module.
      • Image feature extractor: 512-D activation vector contains local visual context information.
      • Localization module returns a confidence map of sound source and a representative visual feature vector z corresponding to location of source of the given input sound.
    • Localization Network: Given extracted visual and sound concepts, the localization networks generate the sound source location.
  • Localizing Sound Source via Listening.
  1. Dataset:
  • Leverage the large unlabeled FlickrSoundNet dataset;
  • Provides annotations for training supervision models.


Strengths / Novelties:

  • By empirically demonstrating the capability of our unsupervised network, we show the model plausibly works in a variety of categories but partially, in that, without prior knowledge, the network can often get to false conclusion.
  • Leveraging small amount of human knowledge can discipline the model, so that it can correct to capture semantically meaningful relationships.
  • Deduce the way of machine understanding about sound source localization in visual scenes.
  • Potential directions for future research, i.e., multi-modal retrieval, sound based saliency or representation learning and its applications.