[세미나] Image-guided Depth Completion: A Non-linear Filters, Convolutions, and Transformer(AMI Lab POSTECH)

💡 본 문서는 'Image-guided Depth Completion: A Non-linear Filters, Convolutions, and Transformer' 영상을 바탕으로 작성한 글이니 참고하시기 바랍니다.

1. Problem

Problem

• Existing Commercial depth sensors: Good
  • Microsoft Kinect
  • Intel RealSense
• Depths points within the same scanning line of LiDAR sensors: Dense
• Still has limitation
  • sensor noise
  • challenging conditions: transparent, shining dark surfaces
  • limited number of scanning lines of LiDAR sensors

How to Solve the Problem? Depth completion

• Spatially propagating the sparse measurements throughout the whole image to get a dense depth prediction
• Goal: completing and reconstructing the whole depth map from sparse depth measurements & corresponding RGB image
• given sparse depth could be highly sparse due to noise or even no measurement being returned from the depth sensor
• it requires depth completion methods to be capable of
  • detecting depth outliers by measuring the spatial relationship between pixels in both local and global perspectives
  • fusing valid depth values from close or even extremely far distance points

local

Non-linear filter: Bilateral filter

• Goal: Image Smoothing
• Split an image into:
  • large-scale features: structure
  • small-scale features: texture

1.1 

1.1.1 

 

참고

• [Youtube] Image-guided Depth Completion: A Non-linear Filters, Convolutions, and Transformer [Kim Kyeongseon]: https://www.youtube.com/watch?v=-tR5rYfin48