Zero‑Barrier AI Development Training: Automatic Driving Scene Image Segmentation with EasyDL

Overview: Baidu EasyDL launches a Zero‑Barrier AI Development Training Camp that offers six live sessions covering industry‑grade projects such as autonomous driving, smart agriculture, video‑stream vehicle‑pedestrian detection, and waste classification. The first lecture demonstrates image‑segmentation model development for autonomous‑driving road scenes using EasyDL.

Course Highlights: The program is zero‑threshold (no algorithm background required), highly practical with multiple industry projects, focuses on a single technical point per project, and provides deployment options that span the entire AI model lifecycle.

Platform Introduction: EasyDL is a dual‑mode AI development platform offering a no‑code automatic training service for enterprise users and a full‑function BML platform for algorithm developers. It supports image, text, audio‑video, and structured‑data training, provides auto‑hyperparameter search, data augmentation, evaluation reports, and multiple deployment options (public cloud, private cloud, edge devices, and integrated hardware‑software solutions).

Model Selection for Autonomous Driving: The lesson explains why image‑segmentation is chosen over object detection for tasks such as vehicle, pedestrian, and tunnel mask extraction. It discusses trade‑offs among model complexity, accuracy, latency, and data‑change cost, recommending low‑complexity models for simple tasks and high‑complexity models for challenging, background‑rich scenarios.

Data Collection & Processing: EasyDL’s EasyData service provides SDKs for fast, cloud‑edge data acquisition, de‑duplication, resolution adjustment, and quality assessment. The workflow emphasizes using real‑world business data for training, balancing class distribution, and splitting datasets into training, validation, and test sets (default 7:2:1). It also covers handling label imbalance and the importance of data cleaning.

Annotation Strategies: The session details manual annotation pitfalls (mis‑labeling, missing labels), guidelines for handling occlusion, and the use of multiple annotation templates (polygons, circles, brushes). It introduces collaborative annotation (team roles, reviewers) and intelligent annotation that iteratively trains a lightweight model to auto‑label unannotated data, reducing manual effort by up to 70%.

Data Augmentation & Synthesis: To address dataset bias, EasyDL offers data synthesis and augmentation techniques that can reduce required data volume by up to 90%, especially useful for rapidly changing SKU images in retail scenarios.

Model Training & Deployment Demo: The instructor walks through creating a dataset (Apollo autonomous‑driving images), uploading labeled/unlabeled data in COCO format, training an image‑segmentation model (achieving ~90% mAP), and deploying it as an offline SDK on a Linux server. The demo shows code‑free SDK usage, sequence‑key activation, and inference on sample images.

Q&A Highlights: Participants asked about data cleaning, dataset sources, medical cell labeling, typical accuracy metrics, manual vs. automated annotation, multi‑person annotation workflow, handling occlusion, and offline SDK usage. The answers emphasized EasyData’s cleaning tools, the flexibility of the platform for various domains (e.g., OCR, traffic‑light detection), and the importance of model evaluation in real‑world deployment.

Future sessions will continue at 8 PM tonight, covering additional industry cases such as wheat detection and counting, further demonstrating EasyDL’s zero‑threshold approach to AI development.