YOLO_V3_Ad_Training - 8BitsCoding/RobotMentor GitHub Wiki


์ฐธ๊ณ ์‚ฌ์ดํŠธ์˜ How to improve object detection๋ฅผ ์ฐธ์กฐํ•˜์ž.

.cfg์˜ ๋งˆ์ง€๋ง‰์ค„์˜ [yolo]์˜ random=1๋กœ ์„ค์ •

[yolo]
mask = 0,1,2
anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326
classes=80
num=9
jitter=.3
ignore_thresh = .5
truth_thresh = 1
random=1        # ์—ฌ๊ธฐ

.cfg์˜ height, width๋ฅผ 32์˜ ๋ฐฐ์ˆ˜๋กœ ๋ณ€๊ฒฝ(๊ธฐ๋ณธ์€ 416์ผ ๊ฒƒ)

[net]
# Testing
# batch=1
# subdivisions=1
# Training
batch=64
subdivisions=8
width=416       # ์—ฌ๊ธฐ
height=416      # ์—ฌ๊ธฐ
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

๋ฐ์ดํ„ฐ ์…‹์— ๋ผ๋ฒจ์ด ์—†๋Š”์ง€ ๋‹ค์‹œ ํ™•์ธํ•œ๋‹ค.

ํ•™์Šต ๋ฐ์ดํ„ฐ ์ค‘ ์ตœ์†Œํ•œ ํ•˜๋‚˜๋ผ๋„ ๋น„์Šทํ•œ ์ด๋ฏธ์ง€๋ฅผ ๋„ฃ๋Š”๋‹ค.

ํ•™์Šต์˜ ํšŸ์ˆ˜๋ฅผ classes * 2000 ์ด์ƒ์œผ๋กœ ํ•œ๋‹ค.

๋งŒ์•ฝ ์ธ์‹ํ•˜์ง€ ์•Š๊ธฐ๋ฅผ ์›ํ•˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ๋Š”๊ฒฝ์šฐ ๋ผ๋ฒจ์„ ๋ถ™์ด์ง€ ์•Š๊ณ  ๋„ฃ๋Š”๋‹ค.

ํ•ด์ƒ๋„๋ฅผ 416x416์œผ๋กœ ๋ณ€๊ฒฝํ•ด๋„ ์ธ์‹์„ ์›ํ•˜๋Š” ์˜ค๋ธŒ์ ํŠธ์˜ ํฌ๊ธฐ๊ฐ€ 16x16๋ฏธ๋งŒ์ด๋ผ๋ฉด,layers = -1, 11์œผ๋กœ stride=4์œผ๋กœ ์„ค์ •ํ•œ๋‹ค.

[upsample]
stride=2            # ์—ฌ๊ธฐ

[route]
layers = -1, 36     # ์—ฌ๊ธฐ

์•„๋ž˜์˜ .cfg ํŒŒ์ผ์„ ์“ฐ๋Š” ๊ฒƒ๋„ ๋‚˜์œ์„ ํƒ์€ ์•„๋‹ˆ๋‹ค.

์ธ์‹์„ ์›ํ•˜๋Š” ์˜ค๋ธŒ์ ํŠธ์˜ ์ขŒ/์šฐ ๋“ฑ์„ ๊ตฌ๋ถ„ํ•˜๊ธธ ์›ํ•œ๋‹ค๋ฉด flip=0์„ ์ถ”๊ฐ€ํ•˜์ž

[net]
# Testing
batch=1
subdivisions=1
# Training
# batch=64
# subdivisions=16
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
flip=0                  # ์—ฌ๊ธฐ

์ž์„ธํ•œ ์‚ฌํ•ญ์€ ์—ญ์‹œ ํ™ˆํŽ˜์ด์ง€์—์„œ ํ™•์ธ!