SegNet and U-Net are created for segmentation problem and EfficientNet is created for classification problem. I have a task and it is saying that train these models on the same dataset and compare results. Is it possible?

  • $\begingroup$ Segmentation is a pixel-wise classification so you can use any classification network with added deconvolution or upsampling layers to upscale to original input resolution. $\endgroup$
    – Brale
    Jun 13, 2020 at 13:22
  • $\begingroup$ I see but it seems little bit not make sense because if I add some deconvolutional layers wouldn't the original Efficient model change? If I add deconvolutional layer it would called like "X Segmentation Model based on EfficientNet", right? $\endgroup$
    – Ugurcan
    Jun 13, 2020 at 13:33
  • $\begingroup$ Yes it would be but the encoder is still EfficientNet. $\endgroup$
    – Brale
    Jun 13, 2020 at 13:48
  • $\begingroup$ Which dataset are they suggesting to train these models on? $\endgroup$
    – nbro
    Jun 13, 2020 at 16:41
  • $\begingroup$ There is no specific dataset. I was asked to compare these 3 models on the same dataset which I will choose. That's the why I am confused and it seems to not make sense for me. Brale is right.I could add deconvolutional layer to EfficientNet but I think it would not be EfficientNet anymore. For example, encoder part of SegNet is VGG-16 without fully connected layer. However, after adding some deconvolutional part and some other approaches it is called SegNet instead of VGG-16. $\endgroup$
    – Ugurcan
    Jun 13, 2020 at 17:59


You must log in to answer this question.