Selected article for: "deep learning and training set"

Author: Yang, Lizhi; Herzi, Ilian; Zakhor, Avideh; Hiremath, Anup; Bazargan, Sahm; Tames-Gadam, Robert
Title: Indoor Query System for the Visually Impaired
  • Cord-id: mjh5d9n5
  • Document date: 2020_8_10
  • ID: mjh5d9n5
    Snippet: Scene query is an important problem for the visually impaired population. While existing systems are able to recognize objects surrounding a person, one of their significant shortcomings is that they typically rely on the phone camera with a finite field of view. Therefore, if the object is situated behind the user, it will go undetected unless the user spins around and takes a series of pictures. The recent introduction of affordable, panoramic cameras solves this problem. In addition, most exi
    Document: Scene query is an important problem for the visually impaired population. While existing systems are able to recognize objects surrounding a person, one of their significant shortcomings is that they typically rely on the phone camera with a finite field of view. Therefore, if the object is situated behind the user, it will go undetected unless the user spins around and takes a series of pictures. The recent introduction of affordable, panoramic cameras solves this problem. In addition, most existing systems report all “significant” objects in a given scene to the user, rather than respond to a specific user-generated query as to where an object located. The recent introduction of text-to-speech and speech recognition capabilities on mobile phones paves the way for such user-generated queries, and for audio response generation to the user. In this paper, we exploit the above advancements to develop a query system for the visually impaired utilizing a panoramic camera and a smartphone. We propose three designs for such a system: the first is a handheld device, and the second and third are wearable backpack and ring. In all three cases, the user interacts with our systems verbally regarding whereabouts of objects of interest. We exploit deep learning methods to train our system to recognize objects of interest. Accuracy of our system for the disjoint test data from the same buildings in the training set is 99%, and for test data from new buildings not present in the training data set is 53%.

    Search related documents:
    Co phrase search for related documents
    • Try single phrases listed below for: 1