0
$\begingroup$

In some of computer vision papers I read that they start off with a bigger sized image and use pooling to reduce dimensionality and train on the image with lower resolution. However, they don't mention how they deal with the size difference. If we want to have pixel wise estimation, then this should be an issue, is it not?

For example in the "Stacked Hourglass" paper [1], they use images of shape (256, 256, 3) but their final output resolution is size (64, 64). This, according to the paper, "does not affect the network’s ability to produce precise joint predictions."

How can we make predictions for the original size if our final output size is 4 times smaller? Do we upsample again after getting our final prediction back to our original size?

[1] https://arxiv.org/abs/1603.06937

$\endgroup$
0
$\begingroup$

Found the answer in the paper itself:

"The same technique as Tompson et al. [15] is used for supervision. A Mean- Squared Error (MSE) loss is applied comparing the predicted heatmap to a ground-truth heatmap consisting of a 2D gaussian (with standard deviation of 1 px) centered on the joint location. To improve performance at high precision thresholds the prediction is offset by a quarter of a pixel in the direction of its next highest neighbor before transforming back to the original coordinate space of the image."

So they transform the prediction back to the original size and achieve really good results by that. Should've read the paper more thorougly.

$\endgroup$

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.