[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6DS2S64YWDS72JMHGHW96V7":3,"skill-same:01M6DS2S64YWDS72JMHGHW96V7":97},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":44,"commitSha":45,"skillDirectory":46,"forksCount":47,"repositoryHtmlUrl":48,"repositoryHomepageUrl":6,"ownerHtmlUrl":49,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":50,"install":51,"resources":64},"01M6DS2S64YWDS72JMHGHW96V7","senior-computer-vision",null,"Computer vision engineering skill for object detection, image segmentation, and visual AI systems. Covers CNN and Vision Transformer architectures, YOLO\u002FFaster R-CNN\u002FDETR detection, Mask R-CNN\u002FSAM segmentation, and production deployment with ONNX\u002FTensorRT. Includes PyTorch, torchvision, Ultralytics, Detectron2, and MMDetection frameworks. Use when building detection pipelines, training custom models, optimizing inference, or deploying vision systems.","zhangzhang-111-i\u002Fclaude-skills111","zhangzhang-111-i","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F251344927?v=4","01M6DRBP4NCM1ZCHTT0FW0SQHD","lixiying","01M6DRBPAMFDJSGQ69XPK9KM9J",4,"MIT","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"Computer Vision","vision","senior computer vision","computer","segmentation","systems","visual","frameworks","models","pipelines","use","detr","image","detection","Production","engineering","deployment","training","pytorch","onnx","2026-11-10 10:12:30",false,"---\nname: \"senior-computer-vision\"\ndescription: Computer vision engineering skill for object detection, image segmentation, and visual AI systems. Covers CNN and Vision Transformer architectures, YOLO\u002FFaster R-CNN\u002FDETR detection, Mask R-CNN\u002FSAM segmentation, and production deployment with ONNX\u002FTensorRT. Includes PyTorch, torchvision, Ultralytics, Detectron2, and MMDetection frameworks. Use when building detection pipelines, training custom models, optimizing inference, or deploying vision systems.\n---\n\n# Senior Computer Vision Engineer\n\nProduction computer vision engineering skill for object detection, image segmentation, and visual AI system deployment.\n\n## Table of Contents\n\n- [Quick Start](#quick-start)\n- [Core Expertise](#core-expertise)\n- [Tech Stack](#tech-stack)\n- [Workflow 1: Object Detection Pipeline](#workflow-1-object-detection-pipeline)\n- [Workflow 2: Model Optimization and Deployment](#workflow-2-model-optimization-and-deployment)\n- [Workflow 3: Custom Dataset Preparation](#workflow-3-custom-dataset-preparation)\n- [Architecture Selection Guide](#architecture-selection-guide)\n- [Reference Documentation](#reference-documentation)\n- [Common Commands](#common-commands)\n\n## Quick Start\n\n```bash\n# Generate training configuration for YOLO or Faster R-CNN\npython scripts\u002Fvision_model_trainer.py models\u002F --task detection --arch yolov8\n\n# Analyze model for optimization opportunities (quantization, pruning)\npython scripts\u002Finference_optimizer.py model.pt --target onnx --benchmark\n\n# Build dataset pipeline with augmentations\npython scripts\u002Fdataset_pipeline_builder.py images\u002F --format coco --augment\n```\n\n## Core Expertise\n\nThis skill provides guidance on:\n\n- **Object Detection**: YOLO family (v5-v11), Faster R-CNN, DETR, RT-DETR\n- **Instance Segmentation**: Mask R-CNN, YOLACT, SOLOv2\n- **Semantic Segmentation**: DeepLabV3+, SegFormer, SAM (Segment Anything)\n- **Image Classification**: ResNet, EfficientNet, Vision Transformers (ViT, DeiT)\n- **Video Analysis**: Object tracking (ByteTrack, SORT), action recognition\n- **3D Vision**: Depth estimation, point cloud processing, NeRF\n- **Production Deployment**: ONNX, TensorRT, OpenVINO, CoreML\n\n## Tech Stack\n\n| Category | Technologies |\n|----------|--------------|\n| Frameworks | PyTorch, torchvision, timm |\n| Detection | Ultralytics (YOLO), Detectron2, MMDetection |\n| Segmentation | segment-anything, mmsegmentation |\n| Optimization | ONNX, TensorRT, OpenVINO, torch.compile |\n| Image Processing | OpenCV, Pillow, albumentations |\n| Annotation | CVAT, Label Studio, Roboflow |\n| Experiment Tracking | MLflow, Weights & Biases |\n| Serving | Triton Inference Server, TorchServe |\n\n## Workflow 1: Object Detection Pipeline\n\nUse this workflow when building an object detection system from scratch.\n\n### Step 1: Define Detection Requirements\n\nAnalyze the detection task requirements:\n\n```\nDetection Requirements Analysis:\n- Target objects: [list specific classes to detect]\n- Real-time requirement: [yes\u002Fno, target FPS]\n- Accuracy priority: [speed vs accuracy trade-off]\n- Deployment target: [cloud GPU, edge device, mobile]\n- Dataset size: [number of images, annotations per class]\n```\n\n### Step 2: Select Detection Architecture\n\nChoose architecture based on requirements:\n\n| Requirement | Recommended Architecture | Why |\n|-------------|-------------------------|-----|\n| Real-time (>30 FPS) | YOLOv8\u002Fv11, RT-DETR | Single-stage, optimized for speed |\n| High accuracy | Faster R-CNN, DINO | Two-stage, better localization |\n| Small objects | YOLO + SAHI, Faster R-CNN + FPN | Multi-scale detection |\n| Edge deployment | YOLOv8n, MobileNetV3-SSD | Lightweight architectures |\n| Transformer-based | DETR, DINO, RT-DETR | End-to-end, no NMS required |\n\n### Step 3: Prepare Dataset\n\nConvert annotations to required format:\n\n```bash\n# COCO format (recommended)\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fimages\u002F \\\n    --annotations data\u002Flabels\u002F \\\n    --format coco \\\n    --split 0.8 0.1 0.1 \\\n    --output data\u002Fcoco\u002F\n\n# Verify dataset\npython -c \"from pycocotools.coco import COCO; coco = COCO('data\u002Fcoco\u002Ftrain.json'); print(f'Images: {len(coco.imgs)}, Categories: {len(coco.cats)}')\"\n```\n\n### Step 4: Configure Training\n\nGenerate training configuration:\n\n```bash\n# For Ultralytics YOLO\npython scripts\u002Fvision_model_trainer.py data\u002Fcoco\u002F \\\n    --task detection \\\n    --arch yolov8m \\\n    --epochs 100 \\\n    --batch 16 \\\n    --imgsz 640 \\\n    --output configs\u002F\n\n# For Detectron2\npython scripts\u002Fvision_model_trainer.py data\u002Fcoco\u002F \\\n    --task detection \\\n    --arch faster_rcnn_R_50_FPN \\\n    --framework detectron2 \\\n    --output configs\u002F\n```\n\n### Step 5: Train and Validate\n\n```bash\n# Ultralytics training\nyolo detect train data=data.yaml model=yolov8m.pt epochs=100 imgsz=640\n\n# Detectron2 training\npython train_net.py --config-file configs\u002Ffaster_rcnn.yaml --num-gpus 1\n\n# Validate on test set\nyolo detect val model=runs\u002Fdetect\u002Ftrain\u002Fweights\u002Fbest.pt data=data.yaml\n```\n\n### Step 6: Evaluate Results\n\nKey metrics to analyze:\n\n| Metric | Target | Description |\n|--------|--------|-------------|\n| mAP@50 | >0.7 | Mean Average Precision at IoU 0.5 |\n| mAP@50:95 | >0.5 | COCO primary metric |\n| Precision | >0.8 | Low false positives |\n| Recall | >0.8 | Low missed detections |\n| Inference time | \u003C33ms | For 30 FPS real-time |\n\n## Workflow 2: Model Optimization and Deployment\n\nUse this workflow when preparing a trained model for production deployment.\n\n### Step 1: Benchmark Baseline Performance\n\n```bash\n# Measure current model performance\npython scripts\u002Finference_optimizer.py model.pt \\\n    --benchmark \\\n    --input-size 640 640 \\\n    --batch-sizes 1 4 8 16 \\\n    --warmup 10 \\\n    --iterations 100\n```\n\nExpected output:\n\n```\nBaseline Performance (PyTorch FP32):\n- Batch 1: 45.2ms (22.1 FPS)\n- Batch 4: 89.4ms (44.7 FPS)\n- Batch 8: 165.3ms (48.4 FPS)\n- Memory: 2.1 GB\n- Parameters: 25.9M\n```\n\n### Step 2: Select Optimization Strategy\n\n| Deployment Target | Optimization Path |\n|-------------------|-------------------|\n| NVIDIA GPU (cloud) | PyTorch → ONNX → TensorRT FP16 |\n| NVIDIA GPU (edge) | PyTorch → TensorRT INT8 |\n| Intel CPU | PyTorch → ONNX → OpenVINO |\n| Apple Silicon | PyTorch → CoreML |\n| Generic CPU | PyTorch → ONNX Runtime |\n| Mobile | PyTorch → TFLite or ONNX Mobile |\n\n### Step 3: Export to ONNX\n\n```bash\n# Export with dynamic batch size\npython scripts\u002Finference_optimizer.py model.pt \\\n    --export onnx \\\n    --input-size 640 640 \\\n    --dynamic-batch \\\n    --simplify \\\n    --output model.onnx\n\n# Verify ONNX model\npython -c \"import onnx; model = onnx.load('model.onnx'); onnx.checker.check_model(model); print('ONNX model valid')\"\n```\n\n### Step 4: Apply Quantization (Optional)\n\nFor INT8 quantization with calibration:\n\n```bash\n# Generate calibration dataset\npython scripts\u002Finference_optimizer.py model.onnx \\\n    --quantize int8 \\\n    --calibration-data data\u002Fcalibration\u002F \\\n    --calibration-samples 500 \\\n    --output model_int8.onnx\n```\n\nQuantization impact analysis:\n\n| Precision | Size | Speed | Accuracy Drop |\n|-----------|------|-------|---------------|\n| FP32 | 100% | 1x | 0% |\n| FP16 | 50% | 1.5-2x | \u003C0.5% |\n| INT8 | 25% | 2-4x | 1-3% |\n\n### Step 5: Convert to Target Runtime\n\n```bash\n# TensorRT (NVIDIA GPU)\ntrtexec --onnx=model.onnx --saveEngine=model.engine --fp16\n\n# OpenVINO (Intel)\nmo --input_model model.onnx --output_dir openvino\u002F\n\n# CoreML (Apple)\npython -c \"import coremltools as ct; model = ct.convert('model.onnx'); model.save('model.mlpackage')\"\n```\n\n### Step 6: Benchmark Optimized Model\n\n```bash\npython scripts\u002Finference_optimizer.py model.engine \\\n    --benchmark \\\n    --runtime tensorrt \\\n    --compare model.pt\n```\n\nExpected speedup:\n\n```\nOptimization Results:\n- Original (PyTorch FP32): 45.2ms\n- Optimized (TensorRT FP16): 12.8ms\n- Speedup: 3.5x\n- Accuracy change: -0.3% mAP\n```\n\n## Workflow 3: Custom Dataset Preparation\n\nUse this workflow when preparing a computer vision dataset for training.\n\n### Step 1: Audit Raw Data\n\n```bash\n# Analyze image dataset\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fraw\u002F \\\n    --analyze \\\n    --output analysis\u002F\n```\n\nAnalysis report includes:\n\n```\nDataset Analysis:\n- Total images: 5,234\n- Image sizes: 640x480 to 4096x3072 (variable)\n- Formats: JPEG (4,891), PNG (343)\n- Corrupted: 12 files\n- Duplicates: 45 pairs\n\nAnnotation Analysis:\n- Format detected: Pascal VOC XML\n- Total annotations: 28,456\n- Classes: 5 (car, person, bicycle, dog, cat)\n- Distribution: car (12,340), person (8,234), bicycle (3,456), dog (2,890), cat (1,536)\n- Empty images: 234\n```\n\n### Step 2: Clean and Validate\n\n```bash\n# Remove corrupted and duplicate images\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fraw\u002F \\\n    --clean \\\n    --remove-corrupted \\\n    --remove-duplicates \\\n    --output data\u002Fcleaned\u002F\n```\n\n### Step 3: Convert Annotation Format\n\n```bash\n# Convert VOC to COCO format\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fcleaned\u002F \\\n    --annotations data\u002Fannotations\u002F \\\n    --input-format voc \\\n    --output-format coco \\\n    --output data\u002Fcoco\u002F\n```\n\nSupported format conversions:\n\n| From | To |\n|------|-----|\n| Pascal VOC XML | COCO JSON |\n| YOLO TXT | COCO JSON |\n| COCO JSON | YOLO TXT |\n| LabelMe JSON | COCO JSON |\n| CVAT XML | COCO JSON |\n\n### Step 4: Apply Augmentations\n\n```bash\n# Generate augmentation config\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fcoco\u002F \\\n    --augment \\\n    --aug-config configs\u002Faugmentation.yaml \\\n    --output data\u002Faugmented\u002F\n```\n\nRecommended augmentations for detection:\n\n```yaml\n# configs\u002Faugmentation.yaml\naugmentations:\n  geometric:\n    - horizontal_flip: { p: 0.5 }\n    - vertical_flip: { p: 0.1 }  # Only if orientation invariant\n    - rotate: { limit: 15, p: 0.3 }\n    - scale: { scale_limit: 0.2, p: 0.5 }\n\n  color:\n    - brightness_contrast: { brightness_limit: 0.2, contrast_limit: 0.2, p: 0.5 }\n    - hue_saturation: { hue_shift_limit: 20, sat_shift_limit: 30, p: 0.3 }\n    - blur: { blur_limit: 3, p: 0.1 }\n\n  advanced:\n    - mosaic: { p: 0.5 }  # YOLO-style mosaic\n    - mixup: { p: 0.1 }   # Image mixing\n    - cutout: { num_holes: 8, max_h_size: 32, max_w_size: 32, p: 0.3 }\n```\n\n### Step 5: Create Train\u002FVal\u002FTest Splits\n\n```bash\npython scripts\u002Fdataset_pipeline_builder.py data\u002Faugmented\u002F \\\n    --split 0.8 0.1 0.1 \\\n    --stratify \\\n    --seed 42 \\\n    --output data\u002Ffinal\u002F\n```\n\nSplit strategy guidelines:\n\n| Dataset Size | Train | Val | Test |\n|--------------|-------|-----|------|\n| \u003C1,000 images | 70% | 15% | 15% |\n| 1,000-10,000 | 80% | 10% | 10% |\n| >10,000 | 90% | 5% | 5% |\n\n### Step 6: Generate Dataset Configuration\n\n```bash\n# For Ultralytics YOLO\npython scripts\u002Fdataset_pipeline_builder.py data\u002Ffinal\u002F \\\n    --generate-config yolo \\\n    --output data.yaml\n\n# For Detectron2\npython scripts\u002Fdataset_pipeline_builder.py data\u002Ffinal\u002F \\\n    --generate-config detectron2 \\\n    --output detectron2_config.py\n```\n\n## Architecture Selection Guide\n\n### Object Detection Architectures\n\n| Architecture | Speed | Accuracy | Best For |\n|--------------|-------|----------|----------|\n| YOLOv8n | 1.2ms | 37.3 mAP | Edge, mobile, real-time |\n| YOLOv8s | 2.1ms | 44.9 mAP | Balanced speed\u002Faccuracy |\n| YOLOv8m | 4.2ms | 50.2 mAP | General purpose |\n| YOLOv8l | 6.8ms | 52.9 mAP | High accuracy |\n| YOLOv8x | 10.1ms | 53.9 mAP | Maximum accuracy |\n| RT-DETR-L | 5.3ms | 53.0 mAP | Transformer, no NMS |\n| Faster R-CNN R50 | 46ms | 40.2 mAP | Two-stage, high quality |\n| DINO-4scale | 85ms | 49.0 mAP | SOTA transformer |\n\n### Segmentation Architectures\n\n| Architecture | Type | Speed | Best For |\n|--------------|------|-------|----------|\n| YOLOv8-seg | Instance | 4.5ms | Real-time instance seg |\n| Mask R-CNN | Instance | 67ms | High-quality masks |\n| SAM | Promptable | 50ms | Zero-shot segmentation |\n| DeepLabV3+ | Semantic | 25ms | Scene parsing |\n| SegFormer | Semantic | 15ms | Efficient semantic seg |\n\n### CNN vs Vision Transformer Trade-offs\n\n| Aspect | CNN (YOLO, R-CNN) | ViT (DETR, DINO) |\n|--------|-------------------|------------------|\n| Training data needed | 1K-10K images | 10K-100K+ images |\n| Training time | Fast | Slow (needs more epochs) |\n| Inference speed | Faster | Slower |\n| Small objects | Good with FPN | Needs multi-scale |\n| Global context | Limited | Excellent |\n| Positional encoding | Implicit | Explicit |\n\n## Reference Documentation\n→ See references\u002Freference-docs-and-commands.md for details\n\n## Performance Targets\n\n| Metric | Real-time | High Accuracy | Edge |\n|--------|-----------|---------------|------|\n| FPS | >30 | >10 | >15 |\n| mAP@50 | >0.6 | >0.8 | >0.5 |\n| Latency P99 | \u003C50ms | \u003C150ms | \u003C100ms |\n| GPU Memory | \u003C4GB | \u003C8GB | \u003C2GB |\n| Model Size | \u003C50MB | \u003C200MB | \u003C20MB |\n\n## Resources\n\n- **Architecture Guide**: `references\u002Fcomputer_vision_architectures.md`\n- **Optimization Guide**: `references\u002Fobject_detection_optimization.md`\n- **Deployment Guide**: `references\u002Fproduction_vision_systems.md`\n- **Scripts**: `scripts\u002F` directory for automation tools\n","# Senior Computer Vision Engineer\n\nProduction computer vision engineering skill for object detection, image segmentation, and visual AI system deployment.\n\n## Table of Contents\n\n- [Quick Start](#quick-start)\n- [Core Expertise](#core-expertise)\n- [Tech Stack](#tech-stack)\n- [Workflow 1: Object Detection Pipeline](#workflow-1-object-detection-pipeline)\n- [Workflow 2: Model Optimization and Deployment](#workflow-2-model-optimization-and-deployment)\n- [Workflow 3: Custom Dataset Preparation](#workflow-3-custom-dataset-preparation)\n- [Architecture Selection Guide](#architecture-selection-guide)\n- [Reference Documentation](#reference-documentation)\n- [Common Commands](#common-commands)\n\n## Quick Start\n\n```bash\n# Generate training configuration for YOLO or Faster R-CNN\npython scripts\u002Fvision_model_trainer.py models\u002F --task detection --arch yolov8\n\n# Analyze model for optimization opportunities (quantization, pruning)\npython scripts\u002Finference_optimizer.py model.pt --target onnx --benchmark\n\n# Build dataset pipeline with augmentations\npython scripts\u002Fdataset_pipeline_builder.py images\u002F --format coco --augment\n```\n\n## Core Expertise\n\nThis skill provides guidance on:\n\n- **Object Detection**: YOLO family (v5-v11), Faster R-CNN, DETR, RT-DETR\n- **Instance Segmentation**: Mask R-CNN, YOLACT, SOLOv2\n- **Semantic Segmentation**: DeepLabV3+, SegFormer, SAM (Segment Anything)\n- **Image Classification**: ResNet, EfficientNet, Vision Transformers (ViT, DeiT)\n- **Video Analysis**: Object tracking (ByteTrack, SORT), action recognition\n- **3D Vision**: Depth estimation, point cloud processing, NeRF\n- **Production Deployment**: ONNX, TensorRT, OpenVINO, CoreML\n\n## Tech Stack\n\n| Category | Technologies |\n|----------|--------------|\n| Frameworks | PyTorch, torchvision, timm |\n| Detection | Ultralytics (YOLO), Detectron2, MMDetection |\n| Segmentation | segment-anything, mmsegmentation |\n| Optimization | ONNX, TensorRT, OpenVINO, torch.compile |\n| Image Processing | OpenCV, Pillow, albumentations |\n| Annotation | CVAT, Label Studio, Roboflow |\n| Experiment Tracking | MLflow, Weights & Biases |\n| Serving | Triton Inference Server, TorchServe |\n\n## Workflow 1: Object Detection Pipeline\n\nUse this workflow when building an object detection system from scratch.\n\n### Step 1: Define Detection Requirements\n\nAnalyze the detection task requirements:\n\n```\nDetection Requirements Analysis:\n- Target objects: [list specific classes to detect]\n- Real-time requirement: [yes\u002Fno, target FPS]\n- Accuracy priority: [speed vs accuracy trade-off]\n- Deployment target: [cloud GPU, edge device, mobile]\n- Dataset size: [number of images, annotations per class]\n```\n\n### Step 2: Select Detection Architecture\n\nChoose architecture based on requirements:\n\n| Requirement | Recommended Architecture | Why |\n|-------------|-------------------------|-----|\n| Real-time (>30 FPS) | YOLOv8\u002Fv11, RT-DETR | Single-stage, optimized for speed |\n| High accuracy | Faster R-CNN, DINO | Two-stage, better localization |\n| Small objects | YOLO + SAHI, Faster R-CNN + FPN | Multi-scale detection |\n| Edge deployment | YOLOv8n, MobileNetV3-SSD | Lightweight architectures |\n| Transformer-based | DETR, DINO, RT-DETR | End-to-end, no NMS required |\n\n### Step 3: Prepare Dataset\n\nConvert annotations to required format:\n\n```bash\n# COCO format (recommended)\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fimages\u002F \\\n    --annotations data\u002Flabels\u002F \\\n    --format coco \\\n    --split 0.8 0.1 0.1 \\\n    --output data\u002Fcoco\u002F\n\n# Verify dataset\npython -c \"from pycocotools.coco import COCO; coco = COCO('data\u002Fcoco\u002Ftrain.json'); print(f'Images: {len(coco.imgs)}, Categories: {len(coco.cats)}')\"\n```\n\n### Step 4: Configure Training\n\nGenerate training configuration:\n\n```bash\n# For Ultralytics YOLO\npython scripts\u002Fvision_model_trainer.py data\u002Fcoco\u002F \\\n    --task detection \\\n    --arch yolov8m \\\n    --epochs 100 \\\n    --batch 16 \\\n    --imgsz 640 \\\n    --output configs\u002F\n\n# For Detectron2\npython scripts\u002Fvision_model_trainer.py data\u002Fcoco\u002F \\\n    --task detection \\\n    --arch faster_rcnn_R_50_FPN \\\n    --framework detectron2 \\\n    --output configs\u002F\n```\n\n### Step 5: Train and Validate\n\n```bash\n# Ultralytics training\nyolo detect train data=data.yaml model=yolov8m.pt epochs=100 imgsz=640\n\n# Detectron2 training\npython train_net.py --config-file configs\u002Ffaster_rcnn.yaml --num-gpus 1\n\n# Validate on test set\nyolo detect val model=runs\u002Fdetect\u002Ftrain\u002Fweights\u002Fbest.pt data=data.yaml\n```\n\n### Step 6: Evaluate Results\n\nKey metrics to analyze:\n\n| Metric | Target | Description |\n|--------|--------|-------------|\n| mAP@50 | >0.7 | Mean Average Precision at IoU 0.5 |\n| mAP@50:95 | >0.5 | COCO primary metric |\n| Precision | >0.8 | Low false positives |\n| Recall | >0.8 | Low missed detections |\n| Inference time | \u003C33ms | For 30 FPS real-time |\n\n## Workflow 2: Model Optimization and Deployment\n\nUse this workflow when preparing a trained model for production deployment.\n\n### Step 1: Benchmark Baseline Performance\n\n```bash\n# Measure current model performance\npython scripts\u002Finference_optimizer.py model.pt \\\n    --benchmark \\\n    --input-size 640 640 \\\n    --batch-sizes 1 4 8 16 \\\n    --warmup 10 \\\n    --iterations 100\n```\n\nExpected output:\n\n```\nBaseline Performance (PyTorch FP32):\n- Batch 1: 45.2ms (22.1 FPS)\n- Batch 4: 89.4ms (44.7 FPS)\n- Batch 8: 165.3ms (48.4 FPS)\n- Memory: 2.1 GB\n- Parameters: 25.9M\n```\n\n### Step 2: Select Optimization Strategy\n\n| Deployment Target | Optimization Path |\n|-------------------|-------------------|\n| NVIDIA GPU (cloud) | PyTorch → ONNX → TensorRT FP16 |\n| NVIDIA GPU (edge) | PyTorch → TensorRT INT8 |\n| Intel CPU | PyTorch → ONNX → OpenVINO |\n| Apple Silicon | PyTorch → CoreML |\n| Generic CPU | PyTorch → ONNX Runtime |\n| Mobile | PyTorch → TFLite or ONNX Mobile |\n\n### Step 3: Export to ONNX\n\n```bash\n# Export with dynamic batch size\npython scripts\u002Finference_optimizer.py model.pt \\\n    --export onnx \\\n    --input-size 640 640 \\\n    --dynamic-batch \\\n    --simplify \\\n    --output model.onnx\n\n# Verify ONNX model\npython -c \"import onnx; model = onnx.load('model.onnx'); onnx.checker.check_model(model); print('ONNX model valid')\"\n```\n\n### Step 4: Apply Quantization (Optional)\n\nFor INT8 quantization with calibration:\n\n```bash\n# Generate calibration dataset\npython scripts\u002Finference_optimizer.py model.onnx \\\n    --quantize int8 \\\n    --calibration-data data\u002Fcalibration\u002F \\\n    --calibration-samples 500 \\\n    --output model_int8.onnx\n```\n\nQuantization impact analysis:\n\n| Precision | Size | Speed | Accuracy Drop |\n|-----------|------|-------|---------------|\n| FP32 | 100% | 1x | 0% |\n| FP16 | 50% | 1.5-2x | \u003C0.5% |\n| INT8 | 25% | 2-4x | 1-3% |\n\n### Step 5: Convert to Target Runtime\n\n```bash\n# TensorRT (NVIDIA GPU)\ntrtexec --onnx=model.onnx --saveEngine=model.engine --fp16\n\n# OpenVINO (Intel)\nmo --input_model model.onnx --output_dir openvino\u002F\n\n# CoreML (Apple)\npython -c \"import coremltools as ct; model = ct.convert('model.onnx'); model.save('model.mlpackage')\"\n```\n\n### Step 6: Benchmark Optimized Model\n\n```bash\npython scripts\u002Finference_optimizer.py model.engine \\\n    --benchmark \\\n    --runtime tensorrt \\\n    --compare model.pt\n```\n\nExpected speedup:\n\n```\nOptimization Results:\n- Original (PyTorch FP32): 45.2ms\n- Optimized (TensorRT FP16): 12.8ms\n- Speedup: 3.5x\n- Accuracy change: -0.3% mAP\n```\n\n## Workflow 3: Custom Dataset Preparation\n\nUse this workflow when preparing a computer vision dataset for training.\n\n### Step 1: Audit Raw Data\n\n```bash\n# Analyze image dataset\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fraw\u002F \\\n    --analyze \\\n    --output analysis\u002F\n```\n\nAnalysis report includes:\n\n```\nDataset Analysis:\n- Total images: 5,234\n- Image sizes: 640x480 to 4096x3072 (variable)\n- Formats: JPEG (4,891), PNG (343)\n- Corrupted: 12 files\n- Duplicates: 45 pairs\n\nAnnotation Analysis:\n- Format detected: Pascal VOC XML\n- Total annotations: 28,456\n- Classes: 5 (car, person, bicycle, dog, cat)\n- Distribution: car (12,340), person (8,234), bicycle (3,456), dog (2,890), cat (1,536)\n- Empty images: 234\n```\n\n### Step 2: Clean and Validate\n\n```bash\n# Remove corrupted and duplicate images\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fraw\u002F \\\n    --clean \\\n    --remove-corrupted \\\n    --remove-duplicates \\\n    --output data\u002Fcleaned\u002F\n```\n\n### Step 3: Convert Annotation Format\n\n```bash\n# Convert VOC to COCO format\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fcleaned\u002F \\\n    --annotations data\u002Fannotations\u002F \\\n    --input-format voc \\\n    --output-format coco \\\n    --output data\u002Fcoco\u002F\n```\n\nSupported format conversions:\n\n| From | To |\n|------|-----|\n| Pascal VOC XML | COCO JSON |\n| YOLO TXT | COCO JSON |\n| COCO JSON | YOLO TXT |\n| LabelMe JSON | COCO JSON |\n| CVAT XML | COCO JSON |\n\n### Step 4: Apply Augmentations\n\n```bash\n# Generate augmentation config\npython scripts\u002Fdataset_pipeline_builder.py data\u002Fcoco\u002F \\\n    --augment \\\n    --aug-config configs\u002Faugmentation.yaml \\\n    --output data\u002Faugmented\u002F\n```\n\nRecommended augmentations for detection:\n\n```yaml\n# configs\u002Faugmentation.yaml\naugmentations:\n  geometric:\n    - horizontal_flip: { p: 0.5 }\n    - vertical_flip: { p: 0.1 }  # Only if orientation invariant\n    - rotate: { limit: 15, p: 0.3 }\n    - scale: { scale_limit: 0.2, p: 0.5 }\n\n  color:\n    - brightness_contrast: { brightness_limit: 0.2, contrast_limit: 0.2, p: 0.5 }\n    - hue_saturation: { hue_shift_limit: 20, sat_shift_limit: 30, p: 0.3 }\n    - blur: { blur_limit: 3, p: 0.1 }\n\n  advanced:\n    - mosaic: { p: 0.5 }  # YOLO-style mosaic\n    - mixup: { p: 0.1 }   # Image mixing\n    - cutout: { num_holes: 8, max_h_size: 32, max_w_size: 32, p: 0.3 }\n```\n\n### Step 5: Create Train\u002FVal\u002FTest Splits\n\n```bash\npython scripts\u002Fdataset_pipeline_builder.py data\u002Faugmented\u002F \\\n    --split 0.8 0.1 0.1 \\\n    --stratify \\\n    --seed 42 \\\n    --output data\u002Ffinal\u002F\n```\n\nSplit strategy guidelines:\n\n| Dataset Size | Train | Val | Test |\n|--------------|-------|-----|------|\n| \u003C1,000 images | 70% | 15% | 15% |\n| 1,000-10,000 | 80% | 10% | 10% |\n| >10,000 | 90% | 5% | 5% |\n\n### Step 6: Generate Dataset Configuration\n\n```bash\n# For Ultralytics YOLO\npython scripts\u002Fdataset_pipeline_builder.py data\u002Ffinal\u002F \\\n    --generate-config yolo \\\n    --output data.yaml\n\n# For Detectron2\npython scripts\u002Fdataset_pipeline_builder.py data\u002Ffinal\u002F \\\n    --generate-config detectron2 \\\n    --output detectron2_config.py\n```\n\n## Architecture Selection Guide\n\n### Object Detection Architectures\n\n| Architecture | Speed | Accuracy | Best For |\n|--------------|-------|----------|----------|\n| YOLOv8n | 1.2ms | 37.3 mAP | Edge, mobile, real-time |\n| YOLOv8s | 2.1ms | 44.9 mAP | Balanced speed\u002Faccuracy |\n| YOLOv8m | 4.2ms | 50.2 mAP | General purpose |\n| YOLOv8l | 6.8ms | 52.9 mAP | High accuracy |\n| YOLOv8x | 10.1ms | 53.9 mAP | Maximum accuracy |\n| RT-DETR-L | 5.3ms | 53.0 mAP | Transformer, no NMS |\n| Faster R-CNN R50 | 46ms | 40.2 mAP | Two-stage, high quality |\n| DINO-4scale | 85ms | 49.0 mAP | SOTA transformer |\n\n### Segmentation Architectures\n\n| Architecture | Type | Speed | Best For |\n|--------------|------|-------|----------|\n| YOLOv8-seg | Instance | 4.5ms | Real-time instance seg |\n| Mask R-CNN | Instance | 67ms | High-quality masks |\n| SAM | Promptable | 50ms | Zero-shot segmentation |\n| DeepLabV3+ | Semantic | 25ms | Scene parsing |\n| SegFormer | Semantic | 15ms | Efficient semantic seg |\n\n### CNN vs Vision Transformer Trade-offs\n\n| Aspect | CNN (YOLO, R-CNN) | ViT (DETR, DINO) |\n|--------|-------------------|------------------|\n| Training data needed | 1K-10K images | 10K-100K+ images |\n| Training time | Fast | Slow (needs more epochs) |\n| Inference speed | Faster | Slower |\n| Small objects | Good with FPN | Needs multi-scale |\n| Global context | Limited | Excellent |\n| Positional encoding | Implicit | Explicit |\n\n## Reference Documentation\n→ See references\u002Freference-docs-and-commands.md for details\n\n## Performance Targets\n\n| Metric | Real-time | High Accuracy | Edge |\n|--------|-----------|---------------|------|\n| FPS | >30 | >10 | >15 |\n| mAP@50 | >0.6 | >0.8 | >0.5 |\n| Latency P99 | \u003C50ms | \u003C150ms | \u003C100ms |\n| GPU Memory | \u003C4GB | \u003C8GB | \u003C2GB |\n| Model Size | \u003C50MB | \u003C200MB | \u003C20MB |\n\n## Resources\n\n- **Architecture Guide**: `references\u002Fcomputer_vision_architectures.md`\n- **Optimization Guide**: `references\u002Fobject_detection_optimization.md`\n- **Deployment Guide**: `references\u002Fproduction_vision_systems.md`\n- **Scripts**: `scripts\u002F` directory for automation tools\n",{"name":5,"description":7},"main","c0a12fc98d85642fa8893e31fc5cf5dfad879de6","engineering-team\u002Fsenior-computer-vision",0,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i","2026-03-13 21:49:18",{"source":8,"skillName":5,"commands":52,"prompt":62,"zipAvailable":63},[53,56,59],{"tool":54,"command":55},"npx","npx skills add zhangzhang-111-i\u002Fclaude-skills111 --skill senior-computer-vision",{"tool":57,"command":58},"bunx","bunx skills add zhangzhang-111-i\u002Fclaude-skills111 --skill senior-computer-vision",{"tool":60,"command":61},"pnpm","pnpm dlx skills add zhangzhang-111-i\u002Fclaude-skills111 --skill senior-computer-vision","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002FSKILL.md",true,[65,71,75,79,83,89,93],{"path":66,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":69,"sourceHtmlUrl":70},"references\u002Fcomputer_vision_architectures.md","REFERENCE","参考文档",16433,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Freferences\u002Fcomputer_vision_architectures.md",{"path":72,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":73,"sourceHtmlUrl":74},"references\u002Fobject_detection_optimization.md",22641,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Freferences\u002Fobject_detection_optimization.md",{"path":76,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":77,"sourceHtmlUrl":78},"references\u002Fproduction_vision_systems.md",34179,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Freferences\u002Fproduction_vision_systems.md",{"path":80,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":81,"sourceHtmlUrl":82},"references\u002Freference-docs-and-commands.md",2602,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Freferences\u002Freference-docs-and-commands.md",{"path":84,"resourceType":85,"resourceTypeDesc":86,"sizeBytes":87,"sourceHtmlUrl":88},"scripts\u002Fdataset_pipeline_builder.py","SCRIPT","脚本",65583,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Fscripts\u002Fdataset_pipeline_builder.py",{"path":90,"resourceType":85,"resourceTypeDesc":86,"sizeBytes":91,"sourceHtmlUrl":92},"scripts\u002Finference_optimizer.py",21386,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Fscripts\u002Finference_optimizer.py",{"path":94,"resourceType":85,"resourceTypeDesc":86,"sizeBytes":95,"sourceHtmlUrl":96},"scripts\u002Fvision_model_trainer.py",21280,"https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fengineering-team\u002Fsenior-computer-vision\u002Fscripts\u002Fvision_model_trainer.py",[98,123,149,180,200,223],{"skillNo":99,"name":100,"nameZh":6,"description":101,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":102,"categories":103,"keywords":104,"publishedAt":39},"01M6DSNYM8PZKEXGYMRX53PVPH","risk-management-specialist","Medical device risk management specialist implementing ISO 14971 throughout product lifecycle. Provides risk analysis, risk evaluation, risk control, and post-production information analysis. Use when user mentions risk management, ISO 14971, risk analysis, FMEA, fault tree analysis, hazard identification, risk control, risk matrix, benefit-risk analysis, residual risk, risk acceptability, or post-market risk.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Frisk-management-specialist\u002FSKILL.md",[],[105,106,107,108,109,110,111,112,33,113,114,115,29,116,117,118,119,120,121,122],"risk","risk management specialist","management","risk management","risk analysis","risk evaluation","product","Evaluation","lifecycle","analysis","user","risk matrix","Market","device","control","iso","medical","iso 14971",{"skillNo":124,"name":125,"nameZh":6,"description":126,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":127,"categories":128,"keywords":129,"publishedAt":39},"01M6DSNSRBGVZWS23TK0ZQ52QC","regulatory-affairs-head","Senior Regulatory Affairs Manager for HealthTech and MedTech companies. Prepares FDA 510(k), De Novo, and PMA submission packages; analyzes regulatory pathways for new medical devices; drafts responses to FDA deficiency letters and Notified Body queries; develops CE marking technical documentation under EU MDR 2017\u002F745; coordinates multi-market approval strategies across FDA, EU, Health Canada, PMDA, and NMPA; and maintains regulatory intelligence on evolving standards. Use when users need to plan or execute FDA submissions, navigate 510(k) or PMA approval processes, achieve CE marking, prepare pre-submission meeting materials, write regulatory strategy documents, respond to agency queries, or manage compliance documentation for medical device market access.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fregulatory-affairs-head\u002FSKILL.md",[],[130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,121,148],"regulatory affairs head","regulatory","head","documentation","documents","packages","planning","devices","plan","access","write","strategy","evidence","standards","manage","intelligence","across","strategies","clinical",{"skillNo":150,"name":151,"nameZh":6,"description":152,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":153,"categories":154,"keywords":165,"publishedAt":39},"01M6DSNN7GNR2HZE66S4D31EP8","quality-manager-qms-iso13485","ISO 13485 Quality Management System implementation and maintenance for medical device organizations. Provides QMS design, documentation control, internal auditing, CAPA management, and certification support. Use when working with medical device quality systems, preparing for ISO 13485 audits, managing regulatory compliance documentation, setting up corrective actions, or building audit preparation programs. Useful for quality management, audit preparation, regulatory compliance, medical device documentation, and corrective action workflows.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-manager-qms-iso13485\u002FSKILL.md",[155,160],{"categoryNo":156,"categoryName":157,"taxonomyType":158,"taxonomyTypeDesc":159},"01JBSEEDCAT000000000000015","视觉设计","SCENARIO","使用场景",{"categoryNo":161,"categoryName":162,"taxonomyType":163,"taxonomyTypeDesc":164},"01JBSEEDCAT000000000000032","设计师","PROFESSION","适用职业",[166,167,168,169,170,171,133,172,173,174,175,176,29,131,177,178,179,119,120,121],"quality manager qms iso13485","quality","design","review","support","workflows","audit","actions","compliance","system","process","up","maintenance","document",{"skillNo":181,"name":182,"nameZh":6,"description":183,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":184,"categories":185,"keywords":189,"publishedAt":39},"01M6DSNE5PQM106PM4EDGPBWGZ","quality-manager-qmr","Senior Quality Manager Responsible Person (QMR) for HealthTech and MedTech companies. Provides quality system governance, management review leadership, regulatory compliance oversight, and quality performance monitoring per ISO 13485 Clause 5.5.2.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-manager-qmr\u002FSKILL.md",[186],{"categoryNo":187,"categoryName":188,"taxonomyType":163,"taxonomyTypeDesc":164},"01JBSEEDCAT000000000000037","律师与法务",[190,167,169,191,192,193,194,174,195,175,107,196,131,197,198,199,120],"quality manager qmr","policy","performance","governance","monitoring","regulatory compliance","oversight","cost","culture","leadership",{"skillNo":201,"name":202,"nameZh":6,"description":203,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":204,"categories":205,"keywords":210,"publishedAt":39},"01M6DSN7MYXB6X42TB1GH2TNSB","quality-documentation-manager","Document control system management for medical device QMS. Covers document numbering, version control, change management, and 21 CFR Part 11 compliance. Use for document control procedures, change control workflow, document numbering, version management, electronic signature compliance, or regulatory documentation review.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fquality-documentation-manager\u002FSKILL.md",[206,209],{"categoryNo":207,"categoryName":208,"taxonomyType":158,"taxonomyTypeDesc":159},"01JBSEEDCAT000000000000002","文档处理",{"categoryNo":187,"categoryName":188,"taxonomyType":163,"taxonomyTypeDesc":164},[133,211,167,169,212,213,172,113,214,175,215,216,29,217,218,219,121,220,221,222],"quality documentation manager","workflow","change","document lifecycle","change management","list","master","system management","documentation review","cfr","21 cfr part 11","change control",{"skillNo":224,"name":225,"nameZh":6,"description":226,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":227,"categories":228,"keywords":229,"publishedAt":39},"01M6DSN3R749H3R20K15QE8X8W","qms-audit-expert","ISO 13485 internal audit expertise for medical device QMS. Covers audit planning, execution, nonconformity classification, and CAPA verification. Use for internal audit planning, audit execution, finding classification, external audit preparation, or audit program management.","https:\u002F\u002Fgithub.com\u002Fzhangzhang-111-i\u002Fclaude-skills111\u002Fblob\u002Fmain\u002Fra-qm-team\u002Fqms-audit-expert\u002FSKILL.md",[],[172,230,231,136,232,233,234,235,107,29,236,118,120,121],"qms audit expert","expert","verification","execution","classification","schedule","checklist"]