Authors: Andreas Meuleman Hyeonjoong Jang Daniel S. Jeon Min H. Kim
Year: 2021
- 4 fisheye cameras to cover 360° FOV
- Binocular stereo model: Binocular Spherical Stereo
- Use spherical coordinates for pixels
- For a pixel in c0 there is a sphere volume in which the pixel can be matched (fig2) depending on the distance
- A good matched is found when the pixel value of the warped pixels in c1 = I_c0
- Selection of the best camera for stereo matching (osef)
- Double sphere model for camera calibration (the double sphere camera model)
- Edge preserving downsampling and upsampling designed for parallel computing
- For depth estimation on blind pixels (occlusion) diffusion kernel based on inpainting direction
- All results performed with GPU
- Quantitative evaluation with a synthetic dataset with blender
Papier intéressant avec pas mal d'astuces sur le filtrage et le traitement des images fish eye. Cependant certaines parties sont dures à comprendre, et ces méthodes sont destinées à la parallélisation