Stereopsis Flashcards
What is stereopsis
Process of recovering 3D geometry from multiple 2D images.
Leverages the slight difference in viewpoints (disparity) between two images to infer depth
Shape from X problem
Perspective cues: vanishing points provide depth information
Texture cues: Objects with detailed texture appear smaller as they recede
Shading Cues: lighting and reflections on surfaces convey depth
Focus cues: changing camera focus reveals the depth of objects
Why multiple views?
Single image -> ambiguous due to forced perspective or occlusions
multiple views help resolve ambiguities thru differing perspectives
Binocular stereo and depth perception
Binocular stereo:
Simulates how two eyes/cams view the same scene from slightly different angles
Difference in disparity is used to calculate depth
Epipolar Geometry
Defines geometric relationship between two images taken from different view points
Epipolar geometry key elements
Epipolar plane: contains 3D point P and the two camera centers
Epipolar lines: where the epipolar plane intersects the image planes
Epipoles: Projection of one camera’s optical center onto the other’s image plane
Epipolar constraint:
Point in one image corresponds to a point along the epipolar line in the other image
Reduces search space for correspondence from 2D to 1D
Epipolar geometry:
- Essential Matrix
encodes epipolar geometry when camera intrinsic parameters are known
links points in camera coordinates
Epipolar geometry:
- Fundamental matrix
Encodes epipolar gemotry for uncalibrated cameras
Maps a point in one image to a line in the other
Depth from disparity
Triangulation:
Calculates the 3D position of a point P from its projection in two images
Depth is inversely proportional to disparity
Small disparities indicate distance objects, while large d indicate nearby objects
Stereo Matching
Goal: Find corresponding points or patches between left and right images
Point matching: match specific features (corners, edges) across images
Region matching: Matches patches instead of individual points to include contextual information
Image rectification
Purpose:
Aligns images so that corresponding points lie on the same horizontal line
Simplifies stereo matching to search along one axis
Method: apply homography to warp images onto a common plane parallel to the baseline.
Applications of Stereopsis
Depth maps: Calc depth for each pixel, used in augmented reality, 3d reconstruction, and robotic navigation
3D reconstruction: recover 3D shapes from stereo images
Visual effects: replays in sports that reconstruct 3D movements
Autonomous vehicles: use depth maps to detect obstacles
Active vs Passive sensing
Passive sensing: uses ambient light and camera pairs, relies on natural textures for matching.
Active sensing: projects patterns or uses lasers to simplify correspondence
- eg. Kinect: projects structured light for depth sensing, LIDAR: emits laser pulses to measure distance
Epipolar in depth
Two cameras:
Image planes
Each camera captures a 2D project of a 3D scene
Optical centers Ol, and OR of left and right cameras
Img plane: Each camera has an image plane where 3D points are projected
