SegmentationAccuracy.m

classdef SegmentationAccuracy < dagnn.Loss

  properties (Transient)
    pixelAccuracy = 0
    meanAccuracy = 0
    meanIntersectionUnion = 0
    confusion = 0
  end

  methods
    function outputs = forward(obj, inputs, params)
      [~,predictions] = max(inputs{1}, [], 3) ;
      predictions = gather(predictions) ;
      labels = gather(inputs{2}) ;

      % compute statistics only on accumulated pixels
      ok = labels > 0 ;
      numPixels = sum(ok(:)) ;
      obj.confusion = obj.confusion + ...
        accumarray([labels(ok),predictions(ok)],1,[21 21]) ;

      % compute various statistics of the confusion matrix
      pos = sum(obj.confusion,2) ;
      res = sum(obj.confusion,1)' ;
      tp = diag(obj.confusion) ;

      obj.pixelAccuracy = sum(tp) / max(1,sum(obj.confusion(:))) ;
      obj.meanAccuracy = mean(tp ./ max(1, pos)) ;
      obj.meanIntersectionUnion = mean(tp ./ max(1, pos + res - tp)) ;

      obj.average = [obj.pixelAccuracy ; obj.meanAccuracy ; obj.meanIntersectionUnion] ;
      obj.numAveraged = obj.numAveraged + numPixels ;
      outputs{1} = obj.average ;
    end

    function [derInputs, derParams] = backward(obj, inputs, params, derOutputs)
      derInputs{1} = [] ;
      derInputs{2} = [] ;
      derParams = {} ;
    end

    function reset(obj)
      obj.confusion = 0 ;
      obj.pixelAccuracy = 0 ;
      obj.meanAccuracy = 0 ;
      obj.meanIntersectionUnion = 0 ;
      obj.average = [0;0;0] ;
      obj.numAveraged = 0 ;
    end

    function str = toString(obj)
      str = sprintf('acc:%.2f, mAcc:%.2f, mIU:%.2f', ...
                    obj.pixelAccuracy, obj.meanAccuracy, obj.meanIntersectionUnion) ;
    end

    function obj = SegmentationAccuracy(varargin)
      obj.load(varargin) ;
    end
  end
end