University of SaskatchewanHARVEST
  • Login
  • Submit Your Work
  • About
    • About HARVEST
    • Guidelines
    • Browse
      • All of HARVEST
      • Communities & Collections
      • By Issue Date
      • Authors
      • Titles
      • Subjects
      • This Collection
      • By Issue Date
      • Authors
      • Titles
      • Subjects
    • My Account
      • Login
      JavaScript is disabled for your browser. Some features of this site may not work without it.
      View Item 
      • HARVEST
      • Electronic Theses and Dissertations
      • Graduate Theses and Dissertations
      • View Item
      • HARVEST
      • Electronic Theses and Dissertations
      • Graduate Theses and Dissertations
      • View Item

      Object Counting with Deep Learning

      Thumbnail
      View/Open
      AICH-THESIS-2019.pdf (28.74Mb)
      Date
      2019-07-04
      Author
      Aich, Shubhra
      Type
      Thesis
      Degree Level
      Masters
      Metadata
      Show full item record
      Abstract
      This thesis explores various empirical aspects of deep learning or convolutional network based models for efficient object counting. First, we train moderately large convolutional networks on comparatively smaller datasets containing few hundred samples from scratch with conventional image processing based data augmentation. Then, we extend this approach for unconstrained, outdoor images using more advanced architectural concepts. Additionally, we propose an efficient, randomized data augmentation strategy based on sub-regional pixel distribution for low-resolution images. Next, the effectiveness of depth-to-space shuffling of feature elements for efficient segmentation is investigated for simpler problems like binary segmentation -- often required in the counting framework. This depth-to-space operation violates the basic assumption of encoder-decoder type of segmentation architectures. Consequently, it helps to train the encoder model as a sparsely connected graph. Nonetheless, we have found comparable accuracy to that of the standard encoder-decoder architectures with our depth-to-space models. After that, the subtleties regarding the lack of localization information in the conventional scalar count loss for one-look models are illustrated. At this point, without using additional annotations, a possible solution is proposed based on the regulation of a network-generated heatmap in the form of a weak, subsidiary loss. The models trained with this auxiliary loss alongside the conventional loss perform much better compared to their baseline counterparts, both qualitatively and quantitatively. Lastly, the intricacies of tiled prediction for high-resolution images are studied in detail, and a simple and effective trick of eliminating the normalization factor in an existing computational block is demonstrated. All of the approaches employed here are thoroughly benchmarked across multiple heterogeneous datasets for object counting against previous, state-of-the-art approaches.
      Degree
      Master of Science (M.Sc.)
      Department
      Computer Science
      Program
      Computer Science
      Supervisor
      Stavness, Ian Kent
      Committee
      Makaroff, Dwight; Roy, Chanchal; Ko, Seok-Bum
      Copyright Date
      April 2019
      URI
      http://hdl.handle.net/10388/12155
      Subject
      Object counting
      Deep learning
      One-look models
      Heatmap Regulation
      Global Sum Pooling
      Collections
      • Graduate Theses and Dissertations
      University of Saskatchewan

      University Library

      © University of Saskatchewan
      Contact Us | Disclaimer | Privacy