From 21e0a60bcf88c435216b0ec83e2146a5a1c8bcf8 Mon Sep 17 00:00:00 2001
From: Jet Hughes <jethughes0@gmail.com>
Date: Tue, 14 Mar 2023 13:14:38 +1300
Subject: [PATCH] vault backup: 2023-03-14 13:14:38

---
 .../05-feature-description-and-matching.md    | 68 +++++++++++++++++++
 content/notes/cosc-342.md                     |  2 +-
 2 files changed, 69 insertions(+), 1 deletion(-)
 create mode 100644 content/notes/05-feature-description-and-matching.md

diff --git a/content/notes/05-feature-description-and-matching.md b/content/notes/05-feature-description-and-matching.md
new file mode 100644
index 000000000..a80669215
--- /dev/null
+++ b/content/notes/05-feature-description-and-matching.md
@@ -0,0 +1,68 @@
+---
+title: "05-feature-description-and-matching"
+tags: 
+- lecture
+- cosc342
+---
+
+last lecure: taking photographs and sticthcing them together. uses feature detection. identity interest points in image sequences. into homography matrix, map from one image into another
+
+![](https://i.imgur.com/I3WL53L.png)
+
+Blob Features
+- corner point dont keep scale
+- an alternative to corner features
+	- have scale as well as location
+- can be created using the "difference of gaussian method"
+	- Two Gaussians of different width 
+	- Subtract wide from narrow 
+	- Bright blobs – high positive response 
+	- Dark blobs – high negative response
+- ![blob detection example corners>blobs](https://i.imgur.com/yOPrJtO.png)
+- ![](https://i.imgur.com/UMcQbrw.png)
+
+FEATURE DESCRIPTION
+- Features are matched on the basis of some descriptor
+- This is usually a list of numbers, represented as a vector
+- Often a high dimensional vector
+- SIFT descriptors, for example, are 128-dimensional 
+- Matching descriptors should be close to each other
+- This distance should be low even if the image changes
+- Translation and rotation in the image plane
+- Changes in viewing angle and distance (and therefore scale)
+- Changes in illumination, brightness, and contrast
+
+>[!INFO] we will use sift descriptors alot. 
+
+A SIMPLE FEATURE DESCRIPTOR
+- Could use pixels near the feature
+- This is easy to do
+- Works well in some cases
+- Example is greyscale, but generalises easily to colour images
+- Take an window ( odd)
+- -dimensional feature vector
+- Compare with Euclidean distance
+- Often easier to use squared distance
+
+Feature invariance
+![](https://i.imgur.com/DTsADVj.png)
+![](https://i.imgur.com/mNxr2XK.png)
+
+SIFT: SCALE-INVARIANT FEATURE TRANSFORM
+- Translation invariance is easy 
+- Scale invariance comes from using Blob features 
+- Descriptor is computed from a window around the feature 
+- The size of the blob determines the size of the window 
+- Brightness invariance comes from using image gradients 
+- Relative brightness of pixels is fairly constant 
+- Rotation invariance by estimating feature orientation 
+- Window is oriented to the dominant image gradient
+
+SIFT FEATURE DETECTION/DESCRIPTION
+- Detect blob features and determine their scale 
+- Compute a Histogram of Gradients around the blob 
+- Peak(s) in the Histogram determine the orientation 
+- A square region is used to compute the descriptor 
+- Square’s size from the blob size; orientation from HoG peaks
+
+
diff --git a/content/notes/cosc-342.md b/content/notes/cosc-342.md
index 887f84d1b..3e4c58c7f 100644
--- a/content/notes/cosc-342.md
+++ b/content/notes/cosc-342.md
@@ -21,7 +21,7 @@ tags:
 - [x] [[02-colour-representation]]
 - [x] [[03-2d-transforms]]
 - [[04-mosaicing]]
-- [[05]]
+- [[05-feature-description-and-matching]]
 
 # Info