Deploy output dimensions

[priyal...@vvdntech.in]

I'm planning to train a model with custom dataset of 70000 images having various height and width. I created the LMDB without resizing. In the train.prototxt i resized the input to 300x300. But i want to perform object detection in a video of 720p(1280x720 size). how should i do this? Do i need to modify something in deploy.prototxt like -

input_shape {
  dim: 1
  dim: 3
  dim: 1280
  dim: 720
}

I don't wish to resize the video. Please help

[Przemek D]

Changing the deploy size is the way to go about this. Note that the dimensions are NCHW, so if you're working with 1280p videos you probably want to do:

input_shape {
  dim: 1
  dim: 3

  dim: 720
  dim: 1280
}

Also note that if you have Crop layers in your network, you will have to recalculate offsets.

[priyal...@vvdntech.in]

Thanks for the reply Przemek, So i only need to make change in deploy.prototxt, and i dont have to make any changes in train.prototxt, right?