7월 프로젝트

left hand, right hand real time detection on Android Using darkflow(used darknet also)

Open the dir 'progress' if want to know about my project more specific.

Yolo+Android (2개월_s)

https://junyoung-jamong.github.io/machine/learning/2019/01/25/Android%EC%97%90%EC%84%9C-%EB%82%B4-YOLO%EB%AA%A8%EB%8D%B8-%EC%82%AC%EC%9A%A9%ED%95%98%EA%B8%B0.html

핸드셋

hand data set 을 활용하는 예제, egohand dataset이 어떤 것이고, label 처리 코드도 존재 (egohands_dataset_clean, egohands_dataset_clean_4classes_ https://github.com/victordibia/handtracking

hand data set 원본 다큐먼트

http://vision.soic.indiana.edu/projects/egohands/

---

1. cfg

# yolov3-mj.cfg
3: batch=1
4: subdivisions=64
6: batch=1
7: subdivisions=64
*17: flip=0*
18: learning_rate=0.01
20: max_batches=4000(class * 2000)
127: filters=21
135: classes=2
171: filters=21
177: classes=2

# yolov2-mj.cfg
3: batch=1
4: subdivisions=64
6: batch=1
7: subdivisions=64
*17: flip=0*
18: learning_rate=0.01
20: max_batches=4000
119: filters=35
125: classes=2

# tiny-yolo-voc-hand.cfg

• If you train the model to distinguish Left and Right objects as separate classes (left/right hand, left/right-turn on road signs, ...) then for disabling flip data augmentation - add flip=0

2. hand.names

left #class 0
right #class 1

3. hand.data

classes = 2
train = data/train.txt
valid = data/test.txt
names = data/hand.names
backup = backup

4. cmd code(darknet)

# train first
darknet.exe detector train data/hand.data cfg/mj-yolov3-tiny.cfg

# train
darknet.exe detector train data/hand.data cfg/yolov3-mj.cfg backup/yolov3-mj_7000.weights

# test
# image
darknet.exe detector test data/hand.data cfg/yolov3-mj.cfg backup/yolov3-mj_7000.weights C:\Users\rhdwb\Desktop\darknet\build\darknet\x64\data\hand\CARDS_COURTYARD_B_T_frame_0036.jpg
# webcam
darknet.exe detector demo data/hand.data cfg/yolov3-mj.cfg backup/yolov3-mj_7000.weights

# darknet dir
cd C:\Users\rhdwb\Desktop\darknet\build\darknet\x64
C:\Users\rhdwb\Desktop\darknet\build\darknet\x64\data\hand\CARDS_COURTYARD_B_T_frame_0036.jpg