Crop a batch of features based on different centers

[Zhihua_Liu]

Hi there.

My model goes to this way:

Read a batch of images and their annotations(represents different object I am interested in different images).

Batch of images (B, C, H, W) feed into a network, output a new feature (B, C, H, W).

Now, how can I center-crop a fixed patch (e.g. 64 pixels)(Ideally should be B, C, 64, 64) from the generated feature maps, but based on different centers.

Can this kind of batch center crop be included within the training step? (My loss take the cropped patch as y_pred compare to some y_true).

[ptrblck]

@Zhihua_Liu

Now, how can I center-crop a fixed patch (e.g. 64 pixels)(Ideally should be B, C, 64, 64) from the generated feature maps, but based on different centers.

You should be able to directly index the tensor and can thus specify the start and end indices (or derive them from the center and the size of the patch).

@Zhihua_Liu

Can this kind of batch center crop be included within the training step?

Yes, slicing a tensor is differentiable, so should also work during training.

[Zhihua_Liu]

Thanks for your answer!

Based on your answer, my init idea is turn the batched annotations into batched indexes, and then batch indexed the feature map to get cropped features. Is that right understanding?

BTW I saw this:

Cropping a minibatch of images, each image a bit differently

I have a tensor named input with dimensions 64x21x21. It is a minibatch of 64 images, each 21x21 pixels. I’d like to crop each image down to 11x11 pixels. So the output tensor I want would have dimensions 64x11x11. I’d like to crop each image around a different “center pixel.” The center pixels are given by a 2-dimensional long tensor named center with dimensions 64x2. For image i, center[i][0] gives the row index and center[i][1] gives the column index for the pixel that should be at the cente…

I think we have the same purpose. I try the code in that reply, however, seems like not a exactly cropping?

Thanks again!