3.1 KiB
3.1 KiB
Pre-Revision Hints
- Write the pseudocode or steps to find the HOG feature.
- What happens if you change the HOG feature from 2x2 to 4x4.
Exam Info
- Don't be misled by previous years.
- Previously 2/3, now 4/5.
- More questions, same time, less time per question.
- Q1 is compulsory.
- More options as there is more to cover.
- Do one of two coding tasks.
- More graphics on 2021/2022 paper than our one.
Exam hints
- Probably won't be asked for algorithms, but could be asked for general steps.
- What are steps when detecting faces, steps to follow to compute HOG of descriptor, steps for camera calibration.
- Very small computations, like how to compute the gradient for a particular pixel in x direction, y direction, etc.
- One could be canvas 2d, one could be threejs, no guarantee that there won't be two on the same technology.
- Can't get away with just studying one technology.
- Theory questions and application questions.
- Doing analysis on an image.
- Won't be too theoretical, going into mathematics, will be more applied.
- One is around camera calibration and the other around feature and descriptors.
- We did basics like calculating differentials.
- These will be the ideas around Q4 and Q5.
- Won't be calculating eigenvalues.
- Basic mathematics.
- Compute the total amount of features you have in HOG.
- Don't focus on solving mathematics for camera calibration, just need to understand how you create an intrinsic matrix given focal length etc.
- Construct K-matrix.
- Anatomy of the camera matrix.
- What the rows of the P-matrix means.
- Feature side:
- Focus on steps need to follow, why you need to.
- Benefit of principal components.
- Just have to do one image analysis question.
- Try not to just pick one, the other one could be difficult.
- Exam won't be easy.
- Make sure you understand concepts covered in class.
- Waqar's part is more novel, so he has given us extra info about exam.
- Nasre's stuff is more overlapping from previous years.
- Will be given useful methods to refer to, like in Sam's papers.
- Answer can be derived from methods given.
- For both Canvas and ThreeJS.
- Definition of surface normal for example, can give formula or describe.
- Nasre tends to be more applied: might ask more applied questions that theoretical questions.
- Analyse this image, write this code.
- Perfect answer for erosion: no equation, description, example.
- We didn't do image de-calibration.
- Hough transform?
- Or adaptive thresholding?
- Think in transforms, draw transforms.
- Nasre really likes transforms.
- It's good to look at previous exam papers.
- Possibly taken question from previous years???
- Was answering question.
- Look at last 3/4 exam papers.
- Wouldn't pass if you only studied last year's paper.
- Try to remember kernel matrices for filters, e.g., smoothing etc.
- Nasre wants us to know the matrix values.
- Can learn how to develop kernel from first principles.
- Here is a one directional filter, make a two directional one.
- like how you make laplacian of gaussian.
- How to convolve filters.