issues
search
KGGS
/
Open-Images-2019-Object-Detection
https://www.kaggle.com/c/open-images-2019-object-detection
0
stars
0
forks
source link
Intro to TF Hub for Object Detection
#2
Open
skydice
opened
4 years ago
skydice
commented
4 years ago
https://www.kaggle.com/xhlulu/intro-to-tf-hub-for-object-detection
skydice
commented
4 years ago
SSD (Single Shot MultiBox Detector)
MobileNetV2
Key Features
Depthwise Separable Convolutions
Prob: conv층에서 매개변수가 너무 많으면 과적합이 될 수 있음
계산량을 줄이면서도 conv층의 표현력을 유지하기 위한 방법
보통의 conv층 연산 input(3
3
3) -> output(1)
depth-wise convolution input(3
3 + 3
3 + 3 * 3) -> output(1 + 1 + 1)
point-wise convolution input(1 + 1 + 1) -> output (1)
전체 연산량에서 8~9배 좋음
Linear Bottlenecks
Prob: Relu 함수의 비선형에 의한 정보 손실로 인해 발생
연산량을 줄이기 위해 1 * 1 conv로 과도하게 압축하다 보니 정보 손실
input manifold를 충분히 담지 못하는 space에서 Relu 함수를 수행하면 정보 손실이 발생
차원수가 충분히 큰 space에서 Relu 함수를 사용하면 정보 손실 보존
마지막 Relu 함수를 지나기 전에 channel expansion
Inverted Residuals
사실 residual block과 비슷
Information Flow Interpretation
표현력과 capacity를 분리 시킴
Reference
https://www.youtube.com/watch?v=mT5Y-Zumbbw&t=460s
https://medium.com/@zurister/depth-wise-convolution-and-depth-wise-separable-convolution-37346565d4ec
https://medium.com/@luis_gonzales/a-look-at-mobilenetv2-inverted-residuals-and-linear-bottlenecks-d49f85c12423
https://m.youtube.com/watch?v=mT5Y-Zumbbw
https://www.slideshare.net/mobile/JinwonLee9/pr132-ssd-single-shot-multibox-detector
https://www.slideshare.net/mobile/JinwonLee9/pr108-mobilenetv2-inverted-residuals-and-linear-bottlenecks