I am trying to wrap my head around multiple output of neural networks, especially output of CNN in image classification with localization:

Lets say we have CNN with 2 cond layers (conv + pool, conv + pool) and 2 fully-connected (FC) layers. At the end, the last FC layer is connected to softmax layer. The softmax layer will output vector of probabilities of classes (let's say we have 4: dog, cat, car, something else). Now my question is, what is the way to retrieve the location of detected object? I mean the vector $(x,y,width,height)$?

If I understand it correctly, I map another layer to the last FC layer (so the last FC layer is connected to softmax layer, and this layer). However how does it outputs the vector I need? I am really confused about this.


Object classification is something very different from object detection/localization. I recommend you having a look at some papers that do object detection using neural networks. The prominent examples are:

All of these networks have publicly available implementations in various neural network frameworks and there are also many blogs explaining them in detail.

  • $\begingroup$ could you please link some blog? All i have found about yolo for example was just few words how it circa works. $\endgroup$
    – Johnyb
    Mar 15 '18 at 18:30
  • $\begingroup$ How about this? $\endgroup$ Mar 15 '18 at 18:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.