Testing Performance

[InstantWindy]

When I test the ADEKChallenge2016 dataset and VOC2012 dataset, I find the result is bad ,how was your test result?

[MathiasGilson]

Hi, have a look at my fix #7

[InstantWindy]

Thanks! I will try ,but I have a question, Why verification set data is resized and will affect the final result,

[MathiasGilson]

During training the CNN learns multiple features from the images it sees. When you use the openCV resize function it deform the image and the features learned during training are not the same so it cannot generalise. As an example if during training the CNN would have consider a small line to be a relevant feature, if the image is deformed (during validation) this line would not be a small line but let say a curved line and thus it would not be identified as a relevant feature by the CNN leading to bad performances

[InstantWindy]

Thank you!, The paper says that input image scale 1, 1/2, 1/4, but your code shows that the input picture scale 1, 1/2 and the 1/4 feature map .you implementation is different from the paper

[MathiasGilson]

that's a question to ask to @aitorzip

[pgr2015]

@InstantWindy Hi I think you misunderstood the paper. Fig. 2 in paper doesn't mean shrinked image will be input directly, watch out the little words in the pic "share weights and computation". This means feature map is actuallly the input.