Computer Vision Fundamentals, Traditional Methods, Deep Learning Advances, and Cloud AI Deployment

This talk introduces the basics of computer vision, explaining how images and videos can be processed to extract high‑level information and describing key sub‑fields such as object recognition, detection, segmentation, motion tracking, 3D reconstruction, visual question answering, and action recognition.

1. Secrets behind viral activities – An example of a WeChat Moments activity that uses face‑matching algorithms to match user photos with historical figures, illustrating the need for a flexible cloud architecture.

2. Computer Vision Basics

2.1 Definition – Computer vision aims to automate visual tasks by extracting abstract information from images and videos, covering tasks like instance recognition, object detection, semantic segmentation, motion tracking, 3D reconstruction, visual QA, and action recognition.

2.2 Imaging – Discusses various image types (aerial, thermal, X‑ray, CT, molecular) and the filters applied to them.

2.3 Processing Levels – Low‑level (denoise, compression, edge detection), mid‑level (classification, segmentation, object detection), high‑level (scene understanding, face recognition, autonomous driving, multimodal problems).

2.4 Target Tracking – Highlights challenges in tracking fast‑moving objects, occlusion, and illumination changes, using NBA video examples.

2.5 Multimodal Problems – Explains the integration of computer vision, NLP, and speech recognition to solve tasks such as visual QA and image‑to‑text generation.

3. Traditional Image‑Processing Methods

3.1 Feature Design – Edge detection, Harris corners, symmetry, scale‑invariant features (SIFT), and HOG descriptors.

3.2 Segmentation & Detection – Watershed algorithm and active‑shape models for object detection.

4. Deep Learning Revolution

4.1 Neural Networks – Describes input, hidden, and output layers, and relates logistic regression and SVM to single‑layer networks.

4.2 Convolutional Neural Networks – Explains convolution, pooling, and fully‑connected layers, and introduces architectures such as Faster‑RCNN and YOLO for object detection.

4.3 AI Application Cases – (1) Face‑matching activity, (2) Face‑fusion, (3) Image‑to‑story generation.

4.4 Cutting‑Edge Vision Research – Discusses LiDAR vs. monocular approaches for autonomous driving and the Orthogonal Feature Transform (OFT) that converts single‑view images to 3D maps.

5. Cloud AI Support

Describes the need for robust cloud infrastructure to serve AI models to millions of users, outlines Tencent Cloud’s solution matrix (face, voice, ASR, TTS, ML platforms, GPU/FPGA servers), and presents product lines such as Huìyǎn, Shéntú, Míngshì, and Mójìng.

Also introduces the private video‑management platform TIMatrix for enterprise video surveillance and analytics.

6. Skill Advancement

Provides visual roadmaps for further learning in computer vision and AI.

Author & Recruitment

Author: Ye Cong, Tencent AI Technical Expert, former Amazon AWS AI manager, with experience in large‑scale cloud systems and AI research.

Recruitment: AI Product Senior Backend Engineer positions in Shenzhen and Beijing, requiring C/C++/Go, database, networking, and AI algorithm experience.