딥러닝 : object detection

Training

Training model

- forwad & backword로 피드백 => 데이터에 맞게 가중치 업데이트

- 경사하상법

 

Optimizer

- 경사하강법 목표는 loss를 최소화

- SGD ( Stochastic Gradient Descent)

   - update gradients per one data

- Momentum

   - velocity term keep going weight's previous gradient direction

- AdaGrad

   - sum of gradient squared

- RMS-prop

   - exponetial moving average

- Adam

   - RMS-prop + Momentum

 

- Adaptive methods(Adam, RMS-prop) is worse to generalize than non-adaptive methods)SGD, Momentim) 

 

Regualization

Regualization

- overfitting을 방지하기 위한 방법

- 손실 함수에 큰 가중치에 불이익

- 모델에 작은 값의 가중치

- L1 Regualization

   - adds a penalty to the error function. Penalty term is the sum of absolute values of weights

- L2 Regualization

   - Penalty term is the sum of squared values of weights

 

Drop out

Drop out

- Drop out some weights randomly in training process

- prevents some weights are biased and has big values

 

Batch normalization

Batch normalization

- 입력 배치 데이터의 평균, 분산에 따라 안정 분포 값이 반환

- make stable input values vefore activation function

---

Object Detection

object classification :  what is the object in an image

object localization : what and where is the single object in an image

object detection :  what and where is the multiple object in an image

 

One stage detection & Two stage dectection

분기 유무 차이

- two-stage : Faster-RCNN

- one-stage : SSD, YOLO

 

One stage Object detetor architecture

Input => Backbone => Neck => Dense Prediction

- Backbone : feature extractor

- Neck : Merge the different resolution deature maps

- Dense Prediction : Predict score of object and bounding box

 

Two stage Object detetor architecture

 1st forward => 2nd forward

- 1st forward : get the object cnadidate regions

- 2nd forward : classify the object in region proposals

 

Grid

- predict the objects in each grid cells

- feaure 맵의 픽셀 수

 

Anchor

- the detector which is predict single bounding box

- predict one object per anchor

- pre-defined bouing box shape

 

Bounding box, objectness score and class score

- objectness score : Object or not

- Class score : cat or dog or car

 

softmax

- 예측한 값의 total 값을 1이 되도록 변경

 

IOU (Intersection over union)

- the metric of how well predicts the bounding box campared with GT box

 

Non-Maximum Suppression

- Filtering the best predicted boxes using IOU and confidence score

 

Prepare data

Object detection dataset

- One image, One GT

- Trainig set / Evaluation set / Test set