Computer Vision
How many cones do we have
6.5M
Is colour camera wide or narrow lens in humans?
Narrow
How many rods do we hav
100M
How many colours can we see
~100million, can only distinguish 10 million shades, and 100 shades of monochrome b&w
What freqs can we see
400-700nm
What freqs cana camera see
350nm to 1000nm
What is the difference in distribution for colour for person v camera
CDC is regular, human is irregular
What is structured light, limitations and advantages
IR dots, limited range, cheap, only good inside
What is TOF, limitations and advantages
Timing based. Accurate depth. Bad Resolution. Cheap. Limited Range
What is Stereo, limitations and advantages
2 Camera. Colour. High Res, Needs Complext Software, Range bounded by separation
What is LIDAR, limitations and advantages
Rotating IR. No IR range limit. v accurate depth. v expensive. low res/ frame rate
What does box filter do
Smooths/ blurs
What does Gaussian filter do
smooths/ blurs
What does linear shift filter do
shifts all pixels
What does Sobel filters do
Finds the gradients at each point
Steps of Canny?
Smooth with Gaussian. Find derivatives with Sobel. Threshold. Apply non-maximal supression to find local maxs.
Effect of gaussian kernel size on canny
big = fine features of image, small = large scale edges
Hough Transform
Converts local edge points to hough space, and each point will represent all shapes thaat can fit through them. voting occurs and the most likely shape wins
Harris Detector
Looks at patch of image. Takes gradients with Sobel. Find the eigenvalues of the autocorrelation matrix for patch of image. If we have 2 big eigenvalues then it is a strong vector. Brightness invariant as based on brightness. Susceptible to deformation and rotation
SIFT
Threshold windows of gradients to get local features. Save gradients in all directions around the good feature. Brightness invarient, rotational invarient, susceptivle to deformation
Steps of deep learning
Input -> convolutional ->nonlinear -> pooling
Precision
#True Positives/(#True Positives + #False Positives)
Recall
#True Positives/(#True positives + #False Negatives)
F1
Combo of Precision and Recall
IOU
Intersection of object and bounding box/ Union of the two
Erosion?
Shrinks object and smooths out thick lines. Makes image smaller
Dilation?
Dilation enlarges object and smooths out holes in image makes image bigger
Close?
Dilation then Erosion makes image same size but fixes errors
Open?
Erosion then Dilation smooths image keeps image same size
Kalman?
Correction finds the actual position of target. Prediction multiplies previous state by a constant. Data association is the constant in this case. If noisy measurement rely on prediction and vice versa
How make kalman smoother
run two filters one forwards one backwards. not real time
Partical filter?
Multiple particles that converge to the true location as the prediction and correction processes occur. Can predict more positions than kalman, and is non-gaussian. More computation needed
Homography?
Transformation that maps point from one plane to another and describes the relationship between two pictures of planar object
Essential Matrix?
Captures the relative geometry between two calibrated cameras observing same 3D object to create a point cloud. Finds relation in pose between two cameras
Bundle Adjustment?
Iterative reduction of a loss function to minimise reprojection error
RANSAC?
Used to fit data to models by random sampling. Eg feature matching pick random subset of features accross two images and calculate homography, then keep going till this minimised
Fiducial Process?
Find marker corners (corner finding + thresholding), calculate homograaphy using corner locations. Check the particular marker ID, and use this to figure out camera pose
Markerless tracking process?
Find all good features accross image. Save information about the info around these to database. Compare features in next frame and figure out the change in pose from this
Ad/dis fiducial vs markerless
Markerless more computationally intensive but dont have to have obnoxious markers everywhere
Gaussian Pyramid?
Progressive blurring (and size reduction) of image which is good for training size invariance
Laplacian Pyramid?
Iteratively subtract the blurred image from the original to gradually blur edges. Good for noise reduction
Wavelet Pyramid?
A bandpassed representation with non-orientated subbands of the image
Steerable Pyramid?
Shows image successively at each scale and orientation good for feature analysis
Visual Odom?
If theotry only need 4 points to estimate camera movement but in theotry need many more - use RANSAC feature tracking. If we know pose at each step we can estimate next pose
SLAM?
Feature detection -> Lucas Kanade -> Pose Estimation - > Mapping of environment (integrate over time) - > Loop closure - update estimate if see same scene twice
Lucas Kanade?
Calculate approx motion of brightness patch. Solve Linear equations to estimate linear displacement vector of window around brightness patch
Ghosting?
Image shows two instances of object due to difference algorithm
Foreground apature?
Hole in object from difference algorithm
Drag flick project
Pose estimation, tracking, CNN
Iris Project
Circular hough, light intensity index, circularity index
Image augmentation
Bilinear interpolation, Gaussian, markov chain, probabalistic diffusion
Cat face detection
Viola jones, fisherface, histogram equalisation
Muscles morphology
Checkerboard, median filtering, canny, open, close, greyscaling
Facemask detection
CNN, fisherface, viola jones, non maximak suppression, IOU
Photo to DXF
Greyscale, gaussian, open, close, thresholding
Books
Canny, binaraisation, OCR, gemini
Wugby
Lucas kanade, object detection CNN, kalman, mean shift
Wockets
Detectron, YOLO, Roboflow
7 segment
Erosion, Dilation, adaptive threshlding, green theorum
Sign language
Colour space conversion, rotation translation, scaling, CNN, random forest
Microfluidics
Erosion, dilation, median blur
Profanity
Greyscale, optical charaacter recognition, gaussian
Volleyball player tracking
Object detection, pose recognition, canny, hough
Pedestrian warning
Instance segmentation, open, close, kalman
Volleyball birds eye
gaussian, canny, dilation, hough, cnn
Photo booth
Viola jones, fisherface, CNN, face encoding
Circuits
Gaussian, Erosion, dilation, thresholding
Eye tracking physios
Circular hough, CNN, Haar cascade
Pancakes
Hough circle, open, close, colour transfer
Tennis tracking
Gaussian, circle hough, dilation, erosion, CNN
List 6 image transformations and distortions
Translation, Rotation, Blurring, Scaling, Perspective Shifting, Shearing, Noise addition
What is visual dynamic range?
(Brightness)
~10^12:1
State the approximate number of centimetres spatial resolution that humans can perceive at 20 metres
0.6CM
Draw RGB diagram
awdawawd
Draw CIE Diagram
asawadfwa
Draw HSV diagram
asdsad
Advantages/ Disadvantages of RGB
Widely used, Simple to understand. Device dependent, Not suitable for every application
Advantages/ dis of HSV
Intuitive for us, Better for artistic work. Conversion required to get it to RGB to display
Ad/ dis of CIE
Device independent, Based on physical environment, Foundantions of colour spaces. Less as intuitive, requires complex math