Introduction to cloud recognition
This article introduces the core principles, expected effects, and differences from planar image tracking of EasyAR CRS (Cloud Recognition Service), helping developers understand applicable scenarios and limitations of cloud recognition.
Basic principles
Cloud Recognition migrates the identification process to the cloud, suitable for scenarios with large target libraries or dynamic updates. The core workflow is as follows:
Image library management
- Create an image library: Upload your target images to the CRS console. The system automatically calculates visual features and adds them as a Target to the backend database.
- CRUD operations: Add, delete, modify, or query targets in your library through the CRS console. After operations, clients can use updates without app redeployment.
Important
Requirements for target images in cloud recognition are identical to planar image tracking. Details: Target image best practices.
Recognition workflow
- Image upload: The client captures the current frame via the camera and sends image data to EasyAR CRS service.
- Cloud matching: The server rapidly searches the cloud target library to match pre-stored Target data (target images uploaded by developers).
- Result return: Upon successful matching, the cloud returns recognition results (Target ID, target image, etc.) to the client, which then displays virtual content or continues subsequent tracking.
Differences from planar image tracking
| Feature | Planar Image Tracking (Local) | Cloud Recognition (Cloud) |
|---|---|---|
| Recognition computation | Performed locally on device | Performed on cloud server |
| Target library size | Limited by memory and loading time, typically ≤100 targets | Up to 100k targets per library, scalable to 100M+ |
| Target updates | Requires app repackaging and redistribution | Real-time upload, immediate effect |
| Network dependency | Not required (offline available) | Mandatory (network needed for recognition requests) |
| Primary function | Recognition + continuous tracking (6DoF pose output) | One-time recognition (target matching) |
Key notes:
- Recognition: Only performs "what is this target" matching without continuous tracking. For tracking, combine with local planar image tracking.
- Applicable scenarios: Large target volumes (e.g., product catalogs, children's books), frequent updates (e.g., event posters), or single-function needs (recognition-only).
Service usage and management
EasyAR CRS provides flexible, secure cloud target management solutions for diverse needs from personal development to enterprise applications.
Library isolation and security
- Multi-library support: Create multiple independent CRS libraries with full isolation. Example:
- Library A: Marketing campaigns, storing product posters.
- Library B: Education, storing textbook illustrations.
- Security mechanism: Each library accessed via unique API Key and Secret, ensuring data security.
Concurrency mode selection
Based on application scale and scan volume, CRS offers two modes:
| Mode | Applicable scenarios | Features | Activation |
|---|---|---|---|
| Basic concurrency | AR app QPS < 50, general scanning | Self-service, stable | Apply via CRS console |
| High concurrency | AR app QPS ≥ 50, high traffic | Dedicated resources, low latency | Contact EasyAR Support for evaluation |
Tip
Startups or testing phases can use basic mode. Upgrade based on actual traffic (e.g., QPS monitoring) after launch.
Library management and API
- Library management: Daily operations (create/delete/upload targets) refer to Library Management.
- CRS API: Comprehensive REST APIs support:
- Health checks: Query service status via API.
- Automation: Batch upload/delete/modify/query targets.
- Utilities: Target recognition scoring, similarity conflict checks.
Note
CRS supports SDK, WeChat Mini Programs, Web integration. SDK requires EasyAR Sense v2.0.0+.
Effects and expected results
Understanding cloud recognition performance helps developers set realistic goals:
Ideal performance
- Fast recognition: <1s latency from capture to result (good network).
- High accuracy: >98% accuracy with clear targets and stable network.
- Large-scale libraries: Up to 100k targets per library.
- Real-time updates: New targets recognizable immediately after upload (network required).
Suboptimal scenarios and solutions
| Phenomenon | Cause | User experience | Solution |
|---|---|---|---|
| High recognition latency | Poor network, slow image upload | Results delayed by seconds | Add user prompts |
| Recognition failure | Blurry image, target not uploaded | No virtual content appears | Verify CRS library status, guide users to stabilize device |
| Target conflict | Excessive similar images in library | Wrong target recognized | Optimize target distinctiveness or split similar images into separate libraries |
Validation methods
- Development phase: Upload test targets to CRS console, validate with HelloARCRS sample, then integrate into your app.
- Testing phase: Test recognition success rate under various conditions: weak network, dynamic updates, scaled libraries.
Best practices
Cloud recognition expands target capacity and enables dynamic updates but sacrifices offline use and real-time tracking. Choose based on needs: small static libraries → local tracking; large dynamic libraries → cloud recognition.
Recommended development workflow:
- Testing phase: Use basic concurrency mode with minimal targets.
- Pre-launch: Estimate concurrency needs, contact support to upgrade high-concurrency mode (1-2 business days).
- Operations phase: Regularly monitor library health via API.
Important
Seasonal traffic alerts: If expecting surges during holidays/events, contact EasyAR support ≥3 business days in advance to avoid service limitations.
For apps needing large libraries + target tracking, combine cloud recognition with planar image tracking. Details: Combining with Planar Image Tracking.