CS231A Midterm Review
May 19, 2017
Midterm Logistics
● In-class midterm at Skilling Auditorium at 3:00-4:20 PM on May 22, 2017
● SCPD students not taking exam at Stanford should coordinate with SCPD.
● Open book and open notes. Not open computer.
● Lectures 1 - 12 (through Image Classification & 2D Object Detection)
● 10 T/F, 10 MC, 3 longer questions (similar to last year’s)
Topics you should know for the Exam
● General knowledge of linear algebra (matrix multiplication, SVD, etc)
● Camera Models and Transformations
● Non-perspective Cameras
● Camera Calibration
● Single View Metrology
● Epipolar Geometry
● Structure from Motion
● Active Stereo and Volumetric Stereo
● Fitting and Matching
● RANSAC
● Hough Transform
● Detectors and Descriptors
● Image Classification
● 2D Object Detection
Homogeneous Coordinates
● Homogenous coordinates allow us to apply a larger variety of
transformations with matrix multiplication
○ For example, we use homogeneous coordinates to handle the 3D -> 2D projection
● Any point (x, y) becomes represented as (x, y, 1)
● More generally (a1, a2, …, an, w) represents the point (a1/w, a2/w, …, an/w)
Types of transformations
● Isometric transformations preserve distances
○ Rotation, translation, reflection
● Similarity transformations preserve shape
○ Rotation, translation, scaling
● Affine transformations preserve parallelism
○ Rotation, translation, scaling, shearing, etc.
○ T(v) = Av + t, A is invertible
● Projective transformations map lines to lines
○ Pretty much everything else
Camera Parameters
● Extrinsic parameters
○ Rotation and translation from the world frame
● Intrinsic parameters
○ Focal length in x and y direction, camera center offset, skew, distortion
○ Most people assume only 5 parameters (for the sake of this class)
Camera Calibration
Camera Calibration
Solve by SVD!
Single View Metrology
● Vanishing points and vanishing lines (horizon)
● This leads to being able to find angles between lines and planes (recall PS1)
● You can also calibrate the camera from a single image!
Epipolar Geometry
● Understanding the geometry of the scene and the cameras
● Should have knowledge of this entire scene and basic triangulation
○ Epipoles, epipolar lines, reprojection error, etc.
Unique Cases of Epipolar Geometry
● Parallel cameras make the epipoles at infinity
● Forward translation make the epipoles in the same location
The Fundamental Matrix
● Relates corresponding points with a single constraint
● 7 degrees of freedom
● Can be found using Eight Point Algorithm and Normalized Eight-Point
algorithm
Solve by SVD!
Structure from Motion
● Estimating both the camera positions and the 3D structure simultaneously
from point correspondences
● You’ve implemented a few algorithms:
○ Factorization method
○ An iterative triangulation method
Active Stereo and Volumetric Stereo
● Active Stereo
○ Replaces one camera with a projector
● Volumetric Stereo
○ Space carving
○ Shadow carving
○ Voxel coloring
RANSAC
● Select random sample of minimum size
● Compute a model from this
● Compute the inliers within the model
● Repeat steps for a fixed amount and return the model with the most inliers
Hough Transform
● Find some parameter space that defines the line, plane, etc. that we’re trying
to estimate
● For each observation, plot in this parameter space
○ Could be points, lines, hyperplanes, etc.
● Grid up the parameter space and find cells with many observations
● If there is a problem, will be similar to the example found in lecture
Detectors and Descriptors
● Corner detectors
○ Harris corner detector
● Edge detectors
○ Find areas of high gradients, but should smooth before doing so to remove noise
○ Should know about Laplacian of Gaussian and Difference of Gaussian
● Blob detection
○ Similar to edge detection, but in 2D
● SIFT
○ A local descriptor around keypoints based on gradients in the image
○ Scale and in-plane rotation invariant
● HOG
○ Implemented in PS3 - you should know about it!
Image Classification and 2D Object Detection
● This questions on the midterm will be purely conceptual - nothing too
difficult
● Bag of words
○ Histogram representation of “words” (features)
● Part of PS4 (don’t need to implement, but skimming the ideas in preparation
for the midterm is a good idea)
○ Sliding window detectors
○ Non-maximal suppression
Exam Advice
● When studying, make a cheat sheet to quickly reference at exam time.
○ Since you have 80 minutes, you do not want to sift through pages of notes.
● You will not need to know the complex math derivations involved in the
course.
○ There are some linear algebra problems, though!
● You will need to know how generally things work and explain them (camera
matrices, SFM, RANSAC, Hough transforms, etc.).
● This review session is not comprehensive of all material on the exam!
