Realtime Video Enhancement via Deep Learning

Accelerated Computing Intelligent Video Analytics General
,
1

Hello Nvidia,

Your work has greatly inspired me, and I truly appreciate your assistance.

Currently, I am using a camera capable of 500fps at a resolution of 800x600 to detect tennis balls. To prevent motion blur, I need to maintain a high shutter speed. However, capturing at 500 fps results in dark images, and utilizing image processing techniques introduces noise into the images.

I am in search of a deep learning pipeline that could mitigate these issues. Do you offer an end-to-end video enhancement pipeline? Additionally, I would like to know if Nvidia provides an optimized toolkit for video enhancement, such as super-resolution or low-light enhancement.

Thank you for your support.

Below are taken with 500fps and 20 fps respectively.

500fps
500fps800×600 217 KB

20fps
20fps800×600 338 KB

5

Seems the exposure time is too short for the camera to capture a human visible image when you set the camera with 500 FPS.

I can see the ball in the picture taken with 500 FPS even it is too dark to be identified. Are you asking for a model which can improve the picture’s quality?

8

The goal is to detect the ball (all four balls in the dark images). What model would you recommend?
(If improving the picture’s quality helps the detection, I would appreciate if a suitable model exists)

9

Unfortunately there is no such model provided by TAO toolkit. You may need to search some image quality improvement algorithm or models by yourself.