I want to ask about MultiScale Object Detection. For the multiscale, if we have an original image do I need to zoom-in at different scales in the image at a particular point or zoom-out at different scales in the image at a particular point?
$\begingroup$
$\endgroup$
asked Nov 17, 2016 at 7:24
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
Multiscale object detection generally consists in smoothing the image with larger and larger kernels, to check which objects persist across different scales. So, zoom-out, mostly. You now can have access to the scale-space representation of the image, as shown below.
One reason is that, without more knowledge, one does not zoom in beyond the pixel width. As illustrated (with humour) in the following picture.
Meanwhile, in certain conditions, subpixel resolution or superresolution is possible.
answered Nov 17, 2016 at 21:44
-
$\begingroup$ For the first image, the size of the image is same its just the filter size is increasing I guess. but is it same if we use MATLAB command imresize and downsample the image from 720x940 to 256x256 and then further 128x128 and 64x64? $\endgroup$ Nov 17, 2016 at 23:24
-
$\begingroup$
imresizeanddownsampleprovide instanced in the large class of scale-space/multiscale techniques. Using the low-pass effect of larger kernels, you can indeed downsample the low-pass images, and obtain something akin to the Gaussian pyramid. YOu can find a description of the above in the survey paper A Panorama on Multiscale Geometric Representations $\endgroup$ Nov 18, 2016 at 12:45