Lecture 5 - Object Detection

Question 1

Intersection over Union

Answer 1

Boxes that overlap a lot = better

Question 2

Bounding Box Regression

Answer 2

Loss function comparing the predicted and ground truth boxes

Question 3

Non-Max Suppression

Answer 3

Avoids repeated detections of the same instance

Time consuming as involves 3 models (CNN, SVM and regression)

Question 4

Fast R-CNN

Answer 4

Generate feature map for whole image
Replace CNN+SVM+Regression with multi task CNN

Question 5

RoI max pooling

Answer 5

Converts the RoI candidates into fixed-size feature map.

Question 6

Fast R-CNN: Location Loss

Answer 6

Lloc(tu,v) = sum(i in {x,y,w,h})(smoothL1(tui-vi))

Equations clearer on slide. v is the ground truth and t is the predicted box

The loss value is 0 when predicted equals actual

Question 7

Key Fast R-CNN Steps

Answer 7

1. Selective search to propose 2000 region candidates for image
2. Apply CNN to gen feature map
3. Extract relative region for each candidate on the map
4. Apply ROI pooling (to -> fixed size map)
5. Map the feature map to object class and location

Question 8

Faster R-CNN

Answer 8

Integrates region proposal algorithm into the CNN ( small additional model)

Question 9

Training faster R-CNN

Answer 9

1. Train RPN (Region Proposal Network)
2. Train FaST R-CNN
3. Fix shared convolutional layers and train RPN
4. Fix … and train Fast R-CNN