Xilinx / xfopencv

Other
329 stars 142 forks source link

Support for Haar/LBP Cascades #1

Open Crazor opened 7 years ago

Crazor commented 7 years ago

Watching yesterday's webinar, "OpenCV on Zynq: Accelerating 4k60 Dense Optical Flow and Stereo Vision", I was encouraged to suggest OpenCV features I'd like to see supported.

I have a project that makes heavy use of LBP cascades for object detection. I had to forego Haar features for performance reasons. LBP performs okay-ish, but accuracy suffers somewhat. Current target hardware is a Samsung Exynos 5422 (quad Cortex A15 2.1GHz, quad Cortex A7 1.5GHz). The system achieves 10-15fps with LBP features at 640x480 (USB camera is capable of 30fps).

One interesting aspect is that training time is very significant (hours for LBP features, days for Haar features, even with rather small windows and basic training sets of 5-20k sampels). I suppose many people would absolutely LOVE to be able to accelerate training, especially with Amazon EC2 F1 instances now available!

varunsanthaseelan commented 7 years ago

@Crazor , thanks a lot for the feedback.

We will consider supporting LBP and Haar features.