4. Feature detectors & descriptors

Designing feature descriptors

Photometric transformations

→ 위치는 동일 but 색깔이 바뀜
→ brightness / contrast

Geometric transformations

→ size, location, rotation 바뀜

Image patch

각 patch의 픽셀값만 사용
→ photometric & geometric 변화가 없다면 괜찮다

Tiny image
다운샘플링 → 아주 작은 변화는 쉽게 거를 수 있음 / 큰 변화도 차이 줄어듦

geometic 바꾸면 적용 불가능

histogram

scale & rotation에 대처가 쉬움

Orientation normalization

$x, y$ : location
$s$ : scale

direction 알면, patch를 rotate해서 대응 가능
noise에 민감할 수 있음

MOPS descriptor

1. given a feature $(x,y,s, \theta)$

2. get 40x40 image patch, subsample every 5th pixel
   → blur & downsampling: 작은 변화에 덜 민감해지도록
   → low frequency filtering, absorbs localization errors

3. subtract mean, divide by standard deviation
   → photometric 변형에 덜 민감해지도록 (색깔 변화)
   → removes bias(brightness) & gain(contrast)

4. Haar Wavelet Transform
   → 간결하게 표현 추출 + 조명 변화 & 다른 변화에 덜 민감해지도록
   → low frequency projection

GIST descriptor

1. compute filter responses (filter bank of Gabor filters)
2. divide image patch into 4x4 cells
3. compute filter response averages for cell
4. size of descriptor is 4x4xN, where N is the size of the filter bank

HOG descriptor

histogram of 'unsigned' gradients
(아까는 색깔 histogram)

2 x 2 x 8

• 2x2: block scale
• 8: 방향의 개수

SIFT

1. multi-scale extrema detection
2. keypoint localization
3. orientation assignment
4. keypoint descriptor ($x, y, s, \theta$)

1. multi-scale extrema detection

Gaussian & DoG (Difference of Gaussian)

2. keypoint localization

3. orientation assignment

4. keypoint descriptor