GlobSIP 16 - Ordered Weighted Averaging Layer

Image Aesthetic Assessment via Deep Semantic Aggregation

Introducion

We proposed a new parameterized pooling layer - Ordered Weighted Averaging Layer to aggregate features from multi-column networks. We first sort features along specific dimension and multiply them with trainable weights to form a aggregated feature. The parameters of the network are trained by end-to-end back-propagation technique. Results on the standard benchmark of aesthetic quality assessment shows the effectiveness of our approach.

Citation

If you find Ordered Weighted Averagin Layer useful in your research, please consider citing:

Image Aesthetic Assessment via Deep Semantic Aggragation
Kung-Hung Lu, Kuang-Yu Chang and Chu-Song Chen
IEEE Global Conference on Signal and Information Processing (GlobalSIP), 2016

Aesthetic quality assessment results:

Performance comparison of different algorithms on AVA dataset. The table shows the accuracy(%) of standard testing set.

Method	AVA Dataset(%)
Murry et al.	68.0
SCNN	71.2
AVG-SCNN	69.9
DCNN	73.3
RDCNN	74.5
AlexNet	72.3
DMA-Net-ImgFu	75.4
Ours	78.6

Installation

Please refer to Caffe prerequisites

How to use ordered weighted layer

Note: we are now just providing CPU version, the GPU version is coming soon...

example :

layer {
  name: "aggregation"
  bottom: "1st_pool5"
  bottom: "2nd_pool5"
  bottom: "3rd_pool5"
  bottom: "4th_pool5"
  type: "OrderedWeighted"
  top: "aggregation"

  ordered_weighted_param {
    OrderOp: DEX              // Sorted in ascending order or descending order
    positive: true            // Force weights to be positive
    axis: 1                   // The axis to aggregate
    weight_filler {           
      type: "gaussian"
      mean: 0.5
      std: 0.1
    }
  }
}

Resources in this paper

comming soon...

Contact

Please feel free to leave suggestions or comments to Kung-Hung Lu ([email protected]), Kuang-Yu Chang ([email protected]) and Chu-Song Chen ([email protected])