mirror of
https://github.com/jackyzha0/quartz.git
synced 2026-03-24 15:05:42 -05:00
89 lines
3.5 KiB
Markdown
89 lines
3.5 KiB
Markdown
---
|
|
title: "342-assignment-01"
|
|
tags:
|
|
- lecture
|
|
- cosc342
|
|
---
|
|
|
|
Choosing images
|
|
- scene type
|
|
- number of image pairs (more than 1)
|
|
- image size
|
|
- image format and compression
|
|
- image geometry
|
|
|
|
use a range of input images
|
|
|
|
# Experiment 1: Feature Matching
|
|
Highly detailed images will have a lot of information such as trees, text, buildings, landscapes etc. Less detailed images with be sparse, these could be bare buildings, walls, images with large blocks of color, etc. Detailed Images will result in a large number of features which are tightly grouped. I predict this will reduce the accuracy of FLANN based matching.
|
|
|
|
HYPOTHESIS:
|
|
FLANN based matching will decrease in accuracy as level of detail increases
|
|
|
|
EXPERIMENT DESIGN:
|
|
|
|
Variables
|
|
- Independent Variables: Level of detail (Number of SIFT features detected)
|
|
- Dependent Variable: Reprojection error of feature matches
|
|
|
|
Procedure
|
|
1. Select a set of image pairs with various levels of detail
|
|
2. For each image pair, detect SIFT features and perform both FLANN and Brute force matching
|
|
3. Calculate the reprojection error of features matched using both methods for each image pair
|
|
4. For each method, plot the accuracy as a function of the number of features/level of detail
|
|
|
|
_control for other factors: lighting, camera settings (exposure, iso, etc), resolution?__
|
|
|
|
**hypothesis super wordy**: fixed
|
|
**what is your error metric**: reprojection error
|
|
**how will you decide statistically if your hypothesis is true or not**: measure correlation between number of features and level of detail
|
|
|
|
RESULTS
|
|
|
|
DISCUSSION
|
|
|
|
|
|
# Experiment 2: RANSAC for Homography Estimation
|
|
QUESTION: How does the choice of RANSAC threshold affect the speed of homography estimation in highly detailed images?
|
|
|
|
HYPOTHESIS: The choice of RANSAC threshold affects the speed of homography estimation in highly detailed images.
|
|
|
|
As the RANSAC threshold increases the speed of homography estimation decreases
|
|
|
|
EXPERIMENT DESIGN:
|
|
|
|
Variables
|
|
- Independent variable: The RANSAC threshold (in pixels).
|
|
- Dependent variable: The time taken to estimate the homography matrix (in seconds).
|
|
|
|
Procedure
|
|
1. Select a set of image pairs with high levels of detail.
|
|
2. For each image pair, estimate the homography matrix that relates one image to the other, using a fixed set of matching features.
|
|
3. Vary the RANSAC threshold from small to large values, and for each value of the threshold, estimate the homography matrix for each image pair, using the same set of matching features.
|
|
4. Measure the time taken to estimate the homography matrix for each value of the threshold and each image pair.
|
|
5. Plot the time taken as a function of the RANSAC threshold, and analyze the results.
|
|
|
|
RESULTS
|
|
|
|
DISCUSSION
|
|
|
|
4. Procedure:
|
|
|
|
|
|
# Todo
|
|
- specify number and type of images to collect
|
|
- choose question
|
|
- develop system for running tests and timing
|
|
- develop system for creating charts
|
|
|
|
# Log
|
|
31/03
|
|
- Collected images from around the lab and out the windows.
|
|
- I created mosaics from the image pairs to see if the images were ok
|
|
- A lot of the images had mismatches in similar areas at the bottom of the match
|
|
- I think this might be because I rotated from my body while taking the photos, rather than rotating just the phone.
|
|
- So I'll try to take some more pictures by rotating just the phone and see how they look
|
|
- I also need to develop a selection of images that I need to get
|
|
- There are also some issues with the display of the mosaic. doesn't always fit perfectly on screen and there were a few instances of line of randomly colours pixels on the screen
|
|
- 
|