3D Vision & Depth MCQ
3D vision introduction, camera calibration, and stereo depth from disparity.
3D Vision Introduction MCQ
3D vision basics
Recovering geometry from images uses cues like stereo disparity, motion parallax, shading, and learning-based monocular depth. Classical stereo relies on calibrated cameras and epipolar rectification to search along scanlines.
Disparity ↔ depth
For parallel rectified cameras, Z ∝ 1/d where d is horizontal disparity—baseline and focal length set the scale.
Building blocks
Projection
3D points map to 2D via intrinsics K and extrinsics [R|t].
Epipolar
Corresponding point lies on a line (1D search after rectification).
Point clouds
Sets of 3D samples from LiDAR, stereo fusion, or depth maps + unprojection.
RGB-D
Registered color + depth enables volumetric and SLAM methods.
Stereo pair
Correspondence → disparity map → depth map → mesh / point cloud
Camera Calibration MCQ
Why calibrate cameras?
Calibration estimates intrinsics (focal length, principal point, skew) and often radial/tangential distortion so that projected rays match real lenses. Planar checkerboard targets (Zhang's method) are standard: each view gives homography constraints that solve for K and distortion, then extrinsics per pose.
Reprojection error
After calibration, compare detected image points with projections of 3D model points; RMS reprojection error should be small (sub-pixel for good setups).
Key ideas
Intrinsic K
Maps normalized camera coordinates to pixel coordinates; includes fx, fy, cx, cy.
Distortion
Brown–Conrady model: k1, k2 radial; p1, p2 tangential before projection.
Zhang's method
Multiple views of a planar pattern; closed-form init then non-linear refinement.
Extrinsics
Per-image R, t from world (target) frame to camera frame.
Calibration pipeline
Capture images → detect corners → estimate K, distortion, poses → optimize jointly → validate error
Stereo Vision MCQ
Stereo vision in practice
Given two calibrated views, find pixel correspondences along epipolar lines (often after rectification), compute disparity, and recover depth via triangulation. Local methods (block matching, SAD/SSD) and global or learning-based matchers trade accuracy, speed, and occlusion handling.
Rectification
Warp both images so epipolar lines are horizontal scanlines—turns 2D search into 1D along rows.
Key ideas
Correspondence
For each pixel in the reference view, find the matching pixel in the second view.
Disparity
Horizontal shift d = x_left − x_right after rectification (sign convention varies).
Cost volume
Stores matching costs over disparities; winner-take-all or optimization (e.g. SGM).
Occlusions
Regions visible in only one image break uniqueness; often detected via consistency checks.
Stereo pipeline
Calibrate → rectify → match → disparity → depth → optional filtering / fusion