Joonghee's Laboratory

https://towardsdatascience.com/what-is-transposed-convolutional-layer-40e5e6e31c11

https://www.google.com/search?q=region+proposal+network&client=ms-android-lgu-kr&hl=en&prmd=niv&sxsrf=ALeKk02TPjkPNu90_mYaKytmxTXZNsAxxA:1586002182706&source=lnms&tbm=isch&sa=X&ved=2ahUKEwjN6tnF3s7oAhVIPHAKHf0qD70Q_AUoAnoECA8QAw&biw=360&bih=612&dpr=3#imgrc=R4u0emrOG4-DsM&imgdii=yX1I5TKC0e9j9M SSD 와는 다르게 feature map에서 output layer 를 하나만 갖는다..

https://3months.tistory.com/436 Loss function으로 사용할수 있다. Log(p)에서 p는 예측(prediction) 의 확률이다. 해당 확률값이 클수록 log(p)는 0 으로 수렴하고 loss function의 값을 0에 가깝게 가져가야한다. log(p) 앞에 곱해지는 값는 실제 확률이다. 즉, 실제 발생가능한 경우에 비례하여 loss function..

1. Batch normalization 2. Weight normalization 3. Layer normalization https://www.slideshare.net/ssuser06e0c5/normalization-72539464 Normalization 방법 by 홍배 김

https://medium.com/@jonathan_hui/ssd-object-detection-single-shot-multibox-detector-for-real-time-processing-9bd8deac0e06

.

https://machinethink.net/blog/mobilenet-v2/

1. Diagnosis 2. Communication among different platorms like classic autosar and others. Ex. SOMEIP 3. Application life cycle based on automotive environment Ex. Execution manager 4. Application portability based on ara api 5.

https://arxiv.org/pdf/1611.08050.pdf I read this paper which Dr. Oh introduced. It doesn't use any anchor. It's based on a kind of heat map. Please, refer to the paper as below. There are two kinds of networks. One is for localization and the other one is for affinity. ttp://blog.daum.net/jhmoon77/17457975

http://m.blog.daum.net/_blog/_m/articleView.do?blogid=0BH6f&articleno=17457982&homedispkind=B0801&daumid=despair52&ISAJAX=true&isUploadFarmUser=true&ISIFRMAE=false&visituserid=UQVg&isRSSConstruct=false&loginfo=registered&vBlogId=0BH6f&isMaster=true&dormancy=false&host=m.blog.daum.net&META_DESC=%EB%8B%A8%EC%A7%80+%EA%B6%81%EA%B8%88%ED%95%9C%EA%B1%B4+%EB%AC%B4%EC%96%B8..