I want to create a fully-convolutional neural net that trains on wider face datasets in order to draw bounding box around faces. The dataset is highly diverse in the image sizes. So my question is, are there no better ways other than to resize the image in order to train a fully convolutional network? Even if there is no other way than to resize the input image what should I do about the annotated bounding boxes? I do not wish to manually annotate all bounding boxes after resizing the images.

  • $\begingroup$ Is it correct that you want to predict box coordinates around faces? Exactly which sort of architecture do you have in mind? $\endgroup$ – Alex Jan 17 at 13:52
  • $\begingroup$ A simple one. Nothing complex. It is going to be a sliding window implemented convolutionally network. It will have a bunch of conv layers, batch norm, max pool, linear layer implemented convolutionally with swish activation. $\endgroup$ – Clown Jan 18 at 12:20
  • $\begingroup$ This is essentialy the first time I am doing a object localization problem. However I am planning to increase it's effectiveness by introducing other things such as deformable conv layers, roi pooling etc. $\endgroup$ – Clown Jan 18 at 12:23

Your Answer

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

Browse other questions tagged or ask your own question.