Computer Vision in Transportation Workshop – Course Overview and Highlights

Overview

The Didi "Computer Vision in Transportation" workshop focuses on practical applications of visual technology in the transportation domain. It covers the underlying methods, advanced research topics, and real‑world use cases.

Course Structure

1. Introduction to Computer Vision – fundamentals such as representation learning, activation functions, neural network architectures, convolution operators, pooling layers, and batch normalization.

2. Advanced Topics – domain adaptation and lightweight models, addressing challenges like model dependence on expensive labeled data, difficulty of transferring models to new environments, high computational resource demands, and deployment constraints on edge devices.

3. Applications – driver identification (face detection & recognition) and driving‑scenario understanding (vision perception, 3D reconstruction, behavior analysis).

Instructors

• Dr. Che Zhengping – Didi AI Labs

• Prof. Guo Yuhong – Chief Researcher, Didi AI Labs & Carleton University

• Dr. Shen Haifeng – Didi AI Labs (face technology)

• Dr. Li Guangyu – Didi AI Labs (driving‑scenario understanding)

Key Topics in the Syllabus

Computer Vision Basics: representation learning, activation functions, network structures, convolutions, pooling, batch normalization.

Image Classification: datasets, roadmap, classification networks, experiments.

Object Detection: region‑based methods, region‑free methods, experiments.

Domain Adaptation: introduction, CV applications, methods.

Lightweight Models: basics, Inception/Xception, SqueezeNet, MobileNet/V2, ShuffleNet/V2.

Driver Identification: application, overview, experiments.

Driving Scenario Understanding: vision perception, 3D reconstruction, behavior analysis.

Highlights

The workshop emphasizes two frontier topics: domain adaptation to enable models to generalize across different environments, and lightweight models that maintain accuracy while reducing size and computational cost.

Practical examples include driver authentication using face detection/recognition and comprehensive driving‑scenario analysis covering pedestrian detection, vehicle detection, lane‑line segmentation, traffic‑light detection, and 3D semantic reconstruction.

Further Reading

Michigan University 40‑page survey on object detection (2020)

Didi AI Labs achieves five world‑first results on WIDER FACE benchmark

Learning Better Features for Face Detection with Feature Fusion and Segmentation Supervision

Virtual Class Enhanced Discriminative Embedding Learning (NeurIPS 2018)