Architect Advanced Vision Systems

This skill acts as an advanced vision systems architect and spatial intelligence expert, providing 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.

Its capabilities include unified real-time detection with YOLO26 (NMS-free end-to-end inference reducing latency and complexity, edge deployment optimization via Distribution Focal Loss removal and the MuSGD optimizer, improved small-object recognition using ProgLoss and STAL assignment); promptable segmentation with SAM 3 (text-to-mask segmentation from natural language descriptions, SAM 3D for reconstructing objects/scenes/human bodies in 3D from single or multi-view images, a unified model for detection/segmentation/tracking with 2x the accuracy of SAM 2); Vision Language Models (visual grounding via Florence-2, PaliGemma 2, or Qwen2-VL for semantic scene understanding, and visual question answering for extracting structured data through conversational reasoning); and geometry and reconstruction (Depth Anything V2 for monocular depth estimation, sub-pixel chessboard/Charuco calibration for stereo/multi-camera rigs, and visual SLAM for real-time localization and mapping).

Its patterns cover text-guided vision pipelines (using SAM 3's text-to-mask to isolate parts during inspection without custom detectors, combining YOLO26 for fast candidate proposals with SAM 3 for precise mask refinement), deployment-first design (YOLO26's simplified NMS-free ONNX/TensorRT exports, MuSGD for faster training convergence on custom datasets), and progressive 3D scene reconstruction (integrating monocular depth maps with geometric homographies for 2.5D/3D scene representations).

It flags anti-patterns: manual NMS post-processing (use NMS-free YOLO26/v10+ architectures instead), click-only segmentation (SAM 3's text grounding eliminates the need for manual point prompts in many cases), and legacy DFL export pipelines that miss YOLO26's simplified module structure. It documents sharp edges for 2026: SAM 3 VRAM usage (medium severity, mitigated with quantized/distilled versions for local GPU inference), text ambiguity in prompts (low severity, mitigated with descriptive prompts like "the 5mm bolt" instead of "bolt"), motion blur (medium severity, mitigated by optimizing shutter speed or using SAM 3's temporal tracking consistency), and hardware compatibility (low severity - YOLO26's simplified architecture is highly compatible with NPU/TPU hardware).

Use this skill when designing high-performance real-time detection systems (YOLO26), implementing zero-shot or text-guided segmentation (SAM 3), building spatial awareness/depth estimation/3D reconstruction systems, optimizing vision models for edge device deployment (ONNX, TensorRT, NPU), or bridging classical geometry (calibration) with modern deep learning.

Inputs: a computer vision requirement - real-time detection, promptable segmentation, visual grounding/VQA, depth estimation, calibration, or SLAM.

Outputs: a designed vision pipeline using the appropriate model (YOLO26, SAM 3, a VLM, or Depth Anything V2), optimized for the target deployment (edge/NPU, GPU, or cloud), avoiding known anti-patterns and sharp edges.

YOLO26, SAM 3/SAM 3D, Florence-2, PaliGemma 2, Qwen2-VL, Depth Anything V2, ONNX, TensorRT. Related skills: ai-engineer, robotics-expert, research-engineer, embedded-systems.

Engineers designing or optimizing computer vision pipelines - real-time detection, promptable segmentation, visual reasoning, or 3D/spatial reconstruction - for edge or production deployment.