Y. Ganin, V. S. Lempitsky. N^4-Fields: Neural Network Nearest Neighbor Fields for Image Transforms. Computing Research Repository, 2014.

Ganin and Lempitsky propose to use a combination of convolutional neural networks and $K$-nereast-neighbor for edge detection. Their implementation is based on the implementation and architecture by Krizhevsky et al. [1], see here, however code is not publicly available. Similar to [2], the convolutional neural network is trained on patches of fixed size. However, as the desired output annotation may have high dimension, principal component analysis (e.g. see [3]) is employed to reduce the target dimensionality. The network is then trained on these new target outputs. Finally, $1$-nearest-neighbor on a subset of the training set is used to annotate new test samples. In practice, Ganin and Lempitsky use a committee of these so called $N^4$-fields (convolutional neural network + $1$-nearest-neighbor) and average across predictions.

  • [1] A. Krizhevsky, I. Sutskever, G. E. Hinton. ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, pages 1097 – 1105, 2012.
  • [2] P. Dollár, C. Zitnick. Structured Forests for Fast Edge Detection. International Conference on Computer Vision, 2013.
  • [3] C. Bishop. Pattern Recognition and Machine Learning. Springer Verlag, New York, 2006.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.