Simple super-resolution on GPU-based atom mobo

Almost 50% of opencv core has now been roughly parallelised for old CUDA GPU's (i.e. low degree of parallelism); about 10% of cv proper is done; and nothing from ml or highgui (the hard one).

This clip shows a simple SR with rough motion estimation converting the sample tree.avi from opencv into a vga-res sequence.

Again, this runs around 1k fps on the atom GPU mobo I have.

The motion estiation is VERY approximate (since these coding tasks are usually completed in 1 night of hacking) -- only a rough "fovea" is tracked across frames and relocated as a whole to improve resolution in the central region of the frame. The region outside the center is only roughly improved (or disimproved if there is too much motion in the peripheral view).

I presume this is something like the working of the human eye.

Category:

Science & Technology

Tags:

• AI
• computer vision
• GPU
• opencv
• super-resolution