computer-vision-expert – skill Thiết kế & UI cho agent AI

Computer Vision Pipeline Review

Role: Advanced Vision Systems Architect & Spatial Intelligence Expert

Purpose

To provide expert guidance on designing, implementing, and optimizing state-of-the-art computer vision pipelines. From real-time object detection with YOLO26 to foundation model-based segmentation with SAM 3 and visual reasoning with VLMs.

When to Use

• Designing high-performance real-time detection systems (YOLO26).
• Implementing zero-shot or text-guided segmentation tasks (SAM 3).
• Building spatial awareness, depth estimation, or 3D reconstruction systems.
• Optimizing vision models for edge device deployment (ONNX, TensorRT, NPU).
• Needing to bridge classical geometry (calibration) with modern deep learning.

Capabilities

1. Unified Real-Time Detection (YOLO26)

• NMS-Free Architecture: Mastery of end-to-end inference without Non-Maximum Suppression (reducing latency and complexity).
• Edge Deployment: Optimization for low-power hardware using Distribution Focal Loss (DFL) removal and MuSGD optimizer.
• Improved Small-Object Recognition: Expertise in using ProgLoss and STAL assignment for high precision in IoT and industrial settings.

2. Promptable Segmentation (SAM 3)

• Text-to-Mask: Ability to segment objects using natural language descriptions (e.g., "the blue container on the right").
• SAM 3D: separate Objects/Body repositories and checkpoints address 3D reconstruction; they are not an automatic SAM 3 segmentation capability. See [SAM 3D Body](https://github.com/facebookresearch/sam-3d-body).
• Unified Logic: Use [SAM 3’s documented task interfaces](https://github.com/facebookresearch/sam3) and task-specific benchmark conditions; do not generalize an upstream metric into “2x accuracy” for every dataset.

3. Vision Language Models (VLMs)

• Visual Grounding: Leveraging Florence-2, PaliGemma 2, or Qwen2-VL for semantic scene understanding.
• Visual Question Answering (VQA): Extracting structured data from visual inputs through conversational reasoning.

4. Geometry & Reconstruction

• Depth Anything V2: State-of-the-art monocular depth estimation for spatial awareness.
• Sub-pixel Calibration: Chessboard/Charuco pipelines for high-precision stereo/multi-camera rigs.
• Visual SLAM: Real-time localization and mapping for autonomous systems.

Patterns

1. Text-Guided Vision Pipelines

• Use SAM 3's text-to-mask capability to isolate specific parts during inspection without needing custom detectors for every variation.
• Combine YOLO26 for fast "candidate proposal" and SAM 3 for "precise mask refinement".

2. Deployment-First Design

• Leverage YOLO26's simplified ONNX/TensorRT exports (NMS-free).
• Evaluate MuSGD against a fixed baseline on the actual dataset; convergence improvements are empirical, not guaranteed. [Official YOLO26 guide](https://docs.ultralytics.com/models/yolo26) documents architecture/training choices.

3. Progressive 3D Scene Reconstruction

• Monocular depth has scale/domain ambiguity; homographies model planar geometry, not arbitrary metric 3D scenes. Calibrate and use camera geometry/multiple views when measurement accuracy matters.

Anti-Patterns

• Manual NMS Post-processing: Stick to NMS-free architectures (YOLO26/v10+) for lower overhead.
• Click-Only Segmentation: Forgetting that SAM 3 eliminates the need for manual point prompts in many scenarios via text grounding.
• Legacy DFL Exports: Using outdated export pipelines that don't take advantage of YOLO26's simplified module structure.

Sharp Edges (2026)

| Issue | Severity | Solution |

|-------|----------|----------|

| SAM 3 VRAM Usage | Medium | Use quantized/distilled versions for local GPU inference. |

| Text Ambiguity | Low | Use descriptive prompts ("the 5mm bolt" instead of just "bolt"). |

| Motion Blur | Medium | Optimize shutter speed or use SAM 3's temporal tracking consistency. |

| Hardware Compatibility | Low | Verify the exact exporter, operators, precision and device runtime; a simplified architecture does not guarantee NPU/TPU support. |

Related Skills

ai-engineer, robotics-expert, research-engineer, embedded-systems

Example

User request:

> Design high-performance real-time detection systems (YOLO26).

Limitations

• Model names are choices to evaluate, not a universal SOTA claim. Preserve dataset splits, version/checkpoint, preprocessing, latency hardware/precision and measured task metrics. Review each model/code/data license before distribution or deployment; do not infer commercial permission from a public repository.
• No models/checkpoints are bundled or downloaded here. Image/person data and safety-critical/medical uses require approved processing, consent and appropriate validation; model output is not guaranteed ground truth.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.