README.md

YOLO(You Only Look Once)!

발표자료 다운로드(pdf)

1. Image detection with Yolov5

✏ Before start

• YOLO v5 git clone

  • https://github.com/ultralytics/yolov5
  • !git clone https://github.com/ultralytics/yolov5.git

• package installation (requirements.txt)

  • %cd /contents/yolov5/
  • !pip install -r requirements.txt

⚙ Preprocessing & training

(1) CCTV dataset

---

• dataset 구성

  • train: 9 internal CCTV images of the mall
  • validation: 9 internal CCTV images in different places from train data (some overlap)

  

• augmentation

  • augment each of the train images by 11 to create 99 train images.
  • validation images are not augmented.

  # augmentation filter
  seq1 = iaa.Affine(scale={'x':(0.5, 1.5), 'y':(0.5, 1.5)}) # 늘리기
  seq2 = iaa.Affine(translate_percent={'x': (-0.2, 0.2), 'y':(-0.2, 0.2)}) # 옆으로 밀기
  seq3 = iaa.Affine(translate_px={"x": (-20, 20), "y": (-20, 20)}) # 위아래로 늘리기
  seq4 = iaa.Affine(rotate=(-45, 45)) # 사진 45도 돌리기
  seq5 = iaa.Affine(shear=(-16, 16)) # 대각선으로 늘리기
  seq7= iaa.Sequential([
                      iaa.Multiply((1.2, 1.5)), 
                      iaa.Fliplr(1.0) 
                      ]) # 밝기 변화 + 좌우반전
  seq8 = iaa.Grayscale(alpha=1.0) # 회색
  seq9 = iaa.Sequential([iaa.Dropout((0.05, 0.1), per_channel=0.5),
                      iaa.Multiply((0.3, 1.5)),
                      iaa.ChannelShuffle(p=1.0)]) # dropout, 픽셀 조정
  seq10 = iaa.GaussianBlur(sigma=1.5) # 흐리게
  seq11 = iaa.Rot90(1) # 90도 회전
  

• test 결과

(2) BLACKPINK dataset

---

• dataset 구성

  • train video:
    • (BLACKPINK Full Cam) lovesick girls full cam https://www.youtube.com/watch?v=Iq6wkVsaCq0
    • BLACKPINK - ‘Lovesick Girls’ 1025 SBS Inkigayo : NO.1 OF THE WEEK https://www.youtube.com/watch?v=MBStYsiE618
  • test video:
    • (BLACKPINK Encore Fancam) | @SBS Inkigayo_2020.10.25. https://www.youtube.com/watch?v=qMsoWTlBCWc

• labeling

  • using YOLO mark

• test 결과

---

2. 참고 논문 리뷰

(1) Faster R-CNN

(2) YOLO v1

(3) YOLO v2

(4) YOLO v3

(5) YOLO v3 Pytorch

• YOLOv3 를 Pytorch 로 구현하는 코드는 이 Github Repo의 코드를 참고했습니다.