Day #260 - Fundamentals Revisited - Detection

Summary of Notes
One stage detector

• Recall high but compromise localization ability, Densebox, Yolo, SSD, RetinaNet
• One Stage Detector - Retinanet
• FPN - Feature Pyramid Net - computing convolutional feature maps
• Class Subnet - object classification 
• box subnet - bounding box regression

Two Stage Detector

• FRCNN, RFCN, MaskRCNN
• Two-stage detector - Strong localization ability
• Feature Pyramid Net Structure
• ROI alignment

Non-Max Suppression - post-processing to eliminate multiple responses
Metrics - Precision, Recall

• Precision - How many selected items were relevant
• Recall - How many relevant items were selected

Detection

• Sliding window approach, Parallel Computation

Manually Handcrafted features for Image

• Haar feature
• Histogram of Gradient
• Local Binary pattern
• Aggregated Channel feature

References 
Link1
Link2
Link3

Happy Mastering Data Science!!!