[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M783VCQ1K9835CKKMF6VFVDT":3,"skill-same:01M783VCQ1K9835CKKMF6VFVDT":76},{"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":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":29,"install":30,"resources":43},"01M783VCQ1K9835CKKMF6VFVDT","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.","motivatedc-creator\u002FSaafy","motivatedc-creator","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F250662783?v=4","01M782H2D8C7XJGXMJC3YHQAWK","Happy Haris","01M782H2ERMBYEM1NSPPPC06S0",0,"MIT","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002FSKILL.md",[],[],"2026-11-15 14:19:17",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\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\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","55ae24ca5523c0499dd70b2874ef674a0e66a85b","engineering-team\u002Fskills\u002Fsenior-computer-vision","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator","2026-07-04 02:45:55",{"source":8,"skillName":5,"commands":31,"prompt":41,"zipAvailable":42},[32,35,38],{"tool":33,"command":34},"npx","npx skills add motivatedc-creator\u002FSaafy --skill senior-computer-vision",{"tool":36,"command":37},"bunx","bunx skills add motivatedc-creator\u002FSaafy --skill senior-computer-vision",{"tool":39,"command":40},"pnpm","pnpm dlx skills add motivatedc-creator\u002FSaafy --skill senior-computer-vision","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002FSKILL.md",true,[44,50,54,58,62,68,72],{"path":45,"resourceType":46,"resourceTypeDesc":47,"sizeBytes":48,"sourceHtmlUrl":49},"references\u002Fcomputer_vision_architectures.md","REFERENCE","参考文档",16433,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Freferences\u002Fcomputer_vision_architectures.md",{"path":51,"resourceType":46,"resourceTypeDesc":47,"sizeBytes":52,"sourceHtmlUrl":53},"references\u002Fobject_detection_optimization.md",22641,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Freferences\u002Fobject_detection_optimization.md",{"path":55,"resourceType":46,"resourceTypeDesc":47,"sizeBytes":56,"sourceHtmlUrl":57},"references\u002Fproduction_vision_systems.md",34179,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Freferences\u002Fproduction_vision_systems.md",{"path":59,"resourceType":46,"resourceTypeDesc":47,"sizeBytes":60,"sourceHtmlUrl":61},"references\u002Freference-docs-and-commands.md",2602,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Freferences\u002Freference-docs-and-commands.md",{"path":63,"resourceType":64,"resourceTypeDesc":65,"sizeBytes":66,"sourceHtmlUrl":67},"scripts\u002Fdataset_pipeline_builder.py","SCRIPT","脚本",65583,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Fscripts\u002Fdataset_pipeline_builder.py",{"path":69,"resourceType":64,"resourceTypeDesc":65,"sizeBytes":70,"sourceHtmlUrl":71},"scripts\u002Finference_optimizer.py",21386,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Fscripts\u002Finference_optimizer.py",{"path":73,"resourceType":64,"resourceTypeDesc":65,"sizeBytes":74,"sourceHtmlUrl":75},"scripts\u002Fvision_model_trainer.py",21280,"https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fengineering-team\u002Fskills\u002Fsenior-computer-vision\u002Fscripts\u002Fvision_model_trainer.py",[77,84,91,98,105,112],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M78558T3RQ22S1B5H0FJAY1Y","syllabus","Generates a curated supplementary reading list from any course syllabus using Consensus academic search. Grill-me intake (syllabus input format + course audience + year range) plus a grouping forcing-options checkpoint before any search runs — so the reading list matches the course's level and recency need. Parses the syllabus to extract topics and learning outcomes, searches Consensus for recent peer-reviewed papers per topic, and produces a professionally formatted .docx with clickable Consensus links, plain-language summaries calibrated to audience level, and Bloom-higher-order discussion questions tied to course learning goals. Use when the user uploads a syllabus, course outline, or curriculum document and wants supplementary readings (e.g., 'create a reading list from this syllabus', 'find recent papers for my course') — even casual mentions with a syllabus attached should trigger this skill.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Fsyllabus\u002Fskills\u002Fsyllabus\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7854YEAAV883M2PKS04TQEK","research","Default entry point for any research request — a hybrid router that classifies the question deterministically and either delegates to a specialist research skill (pulse for trends\u002Fsentiment, grants for NIH funding, litreview for academic literature, syllabus for course reading, patent for prior-art + IP landscape, dossier for entity research) or runs its own plan-decompose-multi-source-search-synthesize-cite fallback workflow when no specialist matches. Always surfaces the routing decision so users can override. Use when the user makes any research request that doesn't obviously match a more-specific specialist skill (e.g., \"research [topic]\", \"look into [topic]\", \"what do we know about [topic]\", \"investigate [topic]\", \"find me information on [topic]\", \"do some research on [topic]\", \"I need to understand [topic]\"). Output is a markdown briefing (default) or .docx document (on request) with full citations and an audit log.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Fresearch\u002Fskills\u002Fresearch\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":19},"01M7854NCGDBQ6Q0QG2YY4VHSX","pulse","Multi-source recency research skill that takes the pulse of any topic across Reddit, Hacker News, the open web, and optionally X\u002FTwitter within a configurable recent window (default 30 days). Forcing intake clarifies topic specificity, angle (trend\u002Fsentiment\u002Fproblems\u002Fopportunities\u002Fcomparison), time window, and platform scope before searching. Returns a synthesized briefing with citations, engagement metrics, and cross-platform pattern analysis. Use when the user requests multi-source recency intelligence on a topic (e.g., 'pulse on [topic]', 'what's happening with [topic]', 'what are people saying about [topic]', 'current conversation about [topic]', 'take the pulse of [topic]', 'trending: [topic]', 'find me info on [topic]'), and for competitor research, trend discovery, tool comparisons, and audience sentiment analysis.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Fpulse\u002Fskills\u002Fpulse\u002FSKILL.md",[],[],{"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":19},"01M78549YZMPYGC1F6S7ZJ4VF8","patent","Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search \u002F freedom-to-operate \u002F competitive landscape \u002F acquisition diligence \u002F litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Use when the user asks for patent searching or analysis (e.g., 'prior art search for [invention]', 'freedom to operate analysis for [product]'). Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Fpatent\u002Fskills\u002Fpatent\u002FSKILL.md",[],[],{"skillNo":106,"name":107,"nameZh":6,"description":108,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":109,"categories":110,"keywords":111,"publishedAt":19},"01M78541PWEQMBPN91PEZZK8H7","notebooklm","Browser automation skill for controlling Google's NotebookLM. Use when the user wants anything done in NotebookLM (e.g., 'open NotebookLM', 'check my [name] notebook', 'ask my notebook about X', 'add [source] to NotebookLM', 'generate a Video Overview from my notebook', 'use NotebookLM Studio'). Handles reading and querying notebooks, adding sources (URLs, text, files, YouTube links, synthesized content), generating Studio outputs (Audio\u002FVideo Overviews, Mind Maps, Reports incl. Briefing Doc\u002FStudy Guide\u002FFAQ, Flashcards, Quiz, slide decks, infographics — discover the exact set from the live Studio panel; the UI evolves fast), and creating new notebooks. Requires browser automation environment — fails gracefully when unavailable.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Fnotebooklm\u002Fskills\u002Fnotebooklm\u002FSKILL.md",[],[],{"skillNo":113,"name":114,"nameZh":6,"description":115,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":116,"categories":117,"keywords":118,"publishedAt":19},"01M7853P7CWJNW5G0Z99YB6H5D","litreview","Academic literature orientation skill that searches papers via free keyless APIs (PubMed E-utilities + OpenAlex) by default — with the Consensus MCP as an optional enhancement lane when connected — builds a strategic search plan using PICO (default) or SPIDER \u002F Decomposition \u002F hybrid as fallbacks, and synthesizes findings into a formatted Word (.docx) research guide. Grill-me intake (research question specificity + framework hint + tentative depth) before the recon search; a second forcing checkpoint after Phase 2 confirms framework + sub-areas + depth before searches consume budget. Configurable depth (5\u002F10\u002F20 queries) controls coverage vs. speed. Output is a 'launching pad' — an orientation guide that lets a researcher dive in confidently, not a finished review. Use when the user starts literature-oriented research (e.g., 'litreview on [topic]', 'literature review on [topic]', 'I'm starting a literature review on X', 'I'm writing a paper on X', 'help me research X', 'I'm doing research on X', 'can you help me research X'). Do NOT use for single one-off paper searches wanting a quick list — that's a plain PubMed\u002FOpenAlex (or Consensus) query.","https:\u002F\u002Fgithub.com\u002Fmotivatedc-creator\u002FSaafy\u002Fblob\u002Fmain\u002Fresearch\u002Flitreview\u002Fskills\u002Flitreview\u002FSKILL.md",[],[]]