System-Design-Notes

[System Design] Street View Blurring

High-level Design

System Responsibility:

• Predicting the location of each object in the image (regression problem)
• Predicting the class of each bounding box (multi-class classification problem)

Two-Stage Networks

• Region proposal network (RPN)
• Classifier

Slower, more accurate

e.g., R-CNN, Fast R-CNN, Faster-RCNN

One-Stage Networks

Use a single network (generate bounding boxes and object classes)

Faster, less accurate

e.g., YOLO, SSD

Feature Engineering

Preprocessing: resizing and normalization

Data augmentation: increase the size of dataset

Two-Stage Object Detection Network

Convolutional layers, Region Proposal Network (RPN), Classifier

Model Training

• Regression loss (e.g., MSE): bounding boxes
• Classification loss (e.g., cross-entropy): classify object
• Final loss: combine the classification loss and regression loss with a parameter

Evaluation

Intersection Over Union (IOU): the overlap between two bounding boxes

Offline metrics:

• precision (varies with different IOU thresholds, for a specific class)
• average precision (various IOU thresholds, for a specific class)
• mean average precision (for all object classes)

Online metrics: user reports and complaints

Overlapping bounding boxes

Non-maximum suppression (NMS): post-algorithm to select the most appropriate box

Summary