Lecture 3 Flashcards
Why is natural feature tracking useful? What are two common features to track?
It avoids the need for targets like checkerboards or markers. Common features to track are corners and blobs.
Why are edges and corners commonly used for feature tracking? How are these formally described?
They are easy to accurately locate, a region with no image gradient is a flat region, a region with a gradient in one direction is an edge, and a gradient in all directions is a corner.
What is a sobel filter? How does this apply to gradient?
A sobel filter is commonly used for edge detection, −1 0 1 −2 0 2 −1 0 1 for horizontal edges, −1 −2 −1 0 0 0 1 2 1 for vertical. The gradient is then a vector which combines both convolutions.
What is Shi Tomasi corner detection?
Works by shifting the image and looking at a region around the original location, if there is a high change for all movements the location is a corner.
what is KLT tracking?
Knade-Lucase-Tomasi tracking, start with Shi-Tomasi corners, we want to find u and v such that old x+u and old y+v equals the next timesteps x and y values. This can be done using linear approximation for 1D.
How can we improve tracking?
KLT tracking only works for small motions because of linear approximations, we could use a pyramid-based optical flow, this works by subsampling the image by half repeatedly, computing the motion at the lowest level, and then doubling to go up a level, this estimate can then be refined. We could also expand on our motion model.
What are blobs?
dark regions surrounded by bright regions, or vice versa.
How can we detect blobs?
Blur the image with larger and larger Gaussian kernels(can be made by repeated blurring with small gaussian kernel, then subtract adjacent images in the stack from one another, blobs will be minima and maxima within these stacks.
How do feature detectors work? Is a simple one feature invariant?
Features are detected on the basis of a descriptor, this is a list of numbers, represented as a vector. The distance between matching vectors should be small, ideally, small regardless of changes in the image. A simple one will be invariant to translation, but not rotation, scale or brightness changes.
What is SIFT features?
An invarient feature detector, scale invariance is done using blob features(descriptor computed from window around feature, size of blob determines window size), Brightness invariance comes from image gradients(gradients do not change much with changing brightness due to relative brightness being the same), rotation invariance is done by finding the dominant gradient direction.
What is space division and approximate neighbours?
to help match SIFT features we split the space into smaller regions, we can then look for neighbours in the same cell as the point we are matching.
What is a k-d tree?
choose an axis and split data along it, trying to split the data roughly in half, then take each half and split again until each cell has only a few items. We can do this for one image and then to find a feature in another image we only have to find the smallest distance feature in that cell.
Why are a lot of SIFT matches wrong? How can we help this?
A lot of blob features don’t have much texture, a lot of scenes have repeating features, to help this we can find the best two matches, keeping the best match only if its distance is much lower than the second bests.
