0%

(CVPR 2017) Deep TEN:Texture Encoding Network

Keyword [DeepTEN]

Zhang H, Xue J, Dana K. Deep ten: Texture encoding network[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 708-717.



1. Overview


In this paper, it proposed Deep Texture Encoding Network (Deep-TEN) with novel Encoding Layer

  • encoding layer. generalizes robust residual encoder


1.1. Contribution

  • encoding layer. learnable residual encoding layer
  • Deep-TEN. feature extraction, dictionary learning and encoding representation learn together



2. Methods


2.1. Residual Encoding Model




  • X. visual descriptor
  • C. learned codebook (learn K x D)
  • r. residual vector
  • E. output
  • S. smoothing factor (learned,

  • Concat K vectors and L2-norm.

    • when K = 1, c = 0, it simplifies to sum pooling
    • can deal with any input size

2.2. Deep-TEN



  • Multi-size Training. 352x352, 320x320
  • pretrained ResNet50
  • K = 32
  • weights of C and s randomly initialized with uniform distribution ±(1/√k)



3. Experiments


3.1. Multi-size vs Single-size training



3.2. Comparison