๋ฐ์ํ
๐ก ๋ณธ ๋ฌธ์๋ 'Image-guided Depth Completion: A Non-linear Filters, Convolutions, and Transformer' ์์์ ๋ฐํ์ผ๋ก ์์ฑํ ๊ธ์ด๋ ์ฐธ๊ณ ํ์๊ธฐ ๋ฐ๋๋๋ค.
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
์ฐธ๊ณ
- [Youtube] Image-guided Depth Completion: A Non-linear Filters, Convolutions, and Transformer [Kim Kyeongseon]: https://www.youtube.com/watch?v=-tR5rYfin48
๋ฐ์ํ