Notice
Recent Posts
Recent Comments
Link
«   2024/05   »
1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31
Archives
Today
Total
관리 메뉴

uos-machine-learning

Yolo v3 논문 리뷰 본문

딥러닝

Yolo v3 논문 리뷰

이산한하루 2019. 9. 18. 00:40

Yolo v3는 이름에서 알 수 있듯이 초기 Yolo Model에서 약간의 디자인의 변화를 준 모델이다. 320 x 320 yolo v3는 22ms 정도로 빠르며, SSD와 거의 정확도가 유사하다는 것을 장점으로 꼽고 있다. 

1. The Deal

1-1. Bounding Box Prediction

network 는 bounding box의 4좌표 tx, ty, tw, th 를 예측한다. 학습하는 동안에는 sum of squared error loss를 사용한다. 그리고 각 bounding box의 objectness socre를 logistic regression 을 이용하여 예측한다. 

1-2. Class Prediction

softmax 함수를 쓰는 것 대신에 독립적인 logistic classifier를 사용한다. 학습하는 동안 binary cross-entropy를 사용한 것을 알 수 있다. 이는 Open Image Dataset을 학습할 때 도움이 된다.

1-3. Predictions Across Scales

Yolo v3는 3가지 다른 크기의 박스를 예측한다. 3개의 feature pyramid network를 사용하여 다양한 feature를 뽑아낸다. 총 9개의 anchor box는 k-means clustering을 통해 결정한다.

1-4. Feature Extractor

1-5. Training

 

 2. Things We Tried Didn't work

Anchor box x, y offset predictions.

Linear x,y predictions instead of logistic.

Focal loss

Dual IOS thresholds and truth assignment.

Comments