2D目标检测

1. requirements

Cloud based or device based deployment

2. ML task & pipeline

3. data collection

4. model

two-stage: region proposal and object classification

generates a set of potential object bounding boxes
takes the proposed regions from the RPN and classifies them into different object categories

one-stage

perform both region proposal and object classification in a single step

nms

5. evaluation

Precision based on IOU threshold
AP: avg. across various IOU thresholds
mAP: mean of AP over C classes

6. deploy & serving

7. 问答

过杀和漏检：基于遗传算法的帕累托优化

Reference