The all-in-one local workbench for computer vision datasets. Annotate, convert, augment, merge, and train — without touching a cloud service or writing a line of code.
Managing CV datasets is a grind. You download a COCO dataset, realise your framework expects YOLO, spend an hour writing a conversion script, discover half your classes are duplicates with different names, and write another script to fix it. Repeat forever.
VisOS wraps all of that in a local UI. No accounts, no uploads, no bill.
Prerequisites: Python 3.10+, Node.js 18+, npm or pnpm
git clone https://github.com/Dan04ggg/VisOS.git
cd VisOS
python3 run.py restartrun.py creates a virtualenv, installs all dependencies, starts the FastAPI backend on :8000 and the Next.js frontend on :3000, health-checks both, and opens your browser.
First run takes 2–5 minutes while PyTorch and Ultralytics download (~1.5 GB).
Load from a local folder or ZIP. Format is auto-detected on load. Datasets persist across restarts via metadata sidecar files.
Supported formats (load & export):
YOLO · COCO · Pascal VOC · LabelMe · CreateML · TensorFlow CSV · ImageNet classification · YOLO OBB · COCO Panoptic · Cityscapes · ADE20K · DOTA · TFRecord
Per-dataset overview: image count, annotation coverage, class distribution chart, train/val/test split breakdown.
Review images one at a time with annotations overlaid. Keyboard-driven — right arrow to keep, left to mark for deletion. Apply bulk changes when done. Filter by annotation status or class. Shift-click for range selection.
Canvas-based editor with six tools:
| Tool | Shortcut |
|---|---|
| Select / Edit | V |
| Bounding Box | B |
| Polygon | P |
| Keypoint | L |
| Brush | R |
| SAM Wand | auto-activates when a SAM model is loaded |
Full undo/redo. Annotations save automatically.
Auto-annotation: load any YOLO, RT-DETR, RF-DETR, SAM, SAM 2/2.1/3, or GroundingDINO model and run inference directly on your dataset with a configurable confidence threshold. GroundingDINO supports zero-shot annotation via text prompt.
Extract, delete, merge, or rename classes without touching JSON. Shows per-class annotation counts.
Convert any supported format to any other. Optionally copy images alongside annotations or annotations only.
Dedicated split view with configurable ratios, optional stratification by class, and a fixed random seed for reproducibility.
Toggle-based pipeline builder. Preview sample outputs before applying. Output to a target image count or a multiplier.
Transforms: horizontal/vertical flip · rotation · scale · translation · shear · perspective · random crop · brightness · contrast · saturation · hue shift · grayscale · Gaussian blur · Gaussian noise · sharpen · JPEG compression · cutout · mosaic · MixUp · elastic deformation · grid distortion · histogram equalisation · channel shuffle · invert · posterize · solarize
Turn video files into annotatable image datasets.
- Every Nth frame
- N frames uniformly distributed across the video
- Keyframes on scene change
- Manual frame selection with scrubber
Supports MP4, AVI, MOV, MKV, WebM.
| Method | What it finds |
|---|---|
| MD5 Hash | Exact byte-for-byte duplicates |
| Perceptual Hash (pHash) | Visually similar images |
| Average Hash (aHash) | Fast approximate similarity |
| CLIP Embeddings | Semantically similar content |
Configurable similarity threshold. Keep strategy: first, largest resolution, or smallest file.
Combine multiple datasets with a class-mapping UI to resolve naming conflicts before merging.
Download pretrained weights from inside the app or import your own .pt, .pth, or .onnx file. Load and unload models to manage GPU memory.
Available pretrained models:
YOLOv5 (n/s) · YOLOv8 (n/s/m/l/x, seg, cls variants) · YOLOv9 (n/s/m/c/e) · YOLOv10 (n/s/m/b/l/x) · RT-DETR (L/X) · RF-DETR (Base/Large) · SAM ViT-B/L · SAM 2 (Tiny/Small/Base+/Large) · SAM 2.1 (Tiny/Small/Base+/Large) · SAM 3 · GroundingDINO (Tiny/Base, zero-shot)
Train locally with live metric monitoring. Supports detection, instance segmentation, and classification tasks.
Architectures: YOLOv8 · YOLOv9 · YOLOv10 · RF-DETR
Configurable: epochs, batch size, image size, learning rate, early-stopping patience
Live: loss, accuracy, validation loss, GPU usage, ETA
Controls: pause, resume, stop with checkpoint saving
Export: PyTorch, ONNX, TensorRT
Track and manage background auto-annotation jobs. Resume interrupted jobs, preview annotated images inline, and monitor per-image progress.
Gallery — grid browser with annotation overlays and full-size click-through
Compare — side-by-side stats between two datasets
Snapshots — save and restore named dataset states before destructive operations
YAML Config — GUI editor for data.yaml files
Health Check — backend API status, Python dependencies, GPU availability, workspace disk usage
python3 run.py start # Start both servers
python3 run.py stop # Stop cleanly
python3 run.py restart # Full restart
python3 run.py restart-back # Backend only
python3 run.py restart-front # Frontend only
python3 run.py status # Show PIDs and ports
python3 run.py logs # Tail live outputNo required environment variables for local use. Backend URL defaults to http://localhost:8000 and is configurable in Settings.
cv-dataset-manager/
├── run.py # Cross-platform process manager
├── backend/
│ ├── main.py # FastAPI entrypoint and all routes
│ ├── dataset_parsers.py # Format auto-detection and parsing
│ ├── format_converter.py # Cross-format conversion
│ ├── annotation_tools.py # Annotation read/write logic
│ ├── augmentation.py # Augmentation pipeline engine
│ ├── dataset_merger.py # Merge with class mapping
│ ├── model_integration.py # Model download, load/unload, inference
│ ├── training.py # Training job management and metric streaming
│ ├── video_utils.py # Frame extraction, duplicate detection, CLIP
│ └── requirements.txt
└── components/ # React views (one per sidebar section)
Proxy pattern: Next.js API routes in app/api/backend/ forward all requests to FastAPI, eliminating CORS issues. The frontend only ever talks to localhost:3000.
Persistence: Datasets survive restarts via dataset_metadata.json sidecars. On startup the backend scans workspace/datasets/ and re-registers everything it finds.
Process manager: run.py handles cross-platform PID tracking, port cleanup, and log streaming — no Docker required.
Base URL: http://localhost:8000/api
Interactive docs: http://localhost:8000/docs
| Resource | Endpoints |
|---|---|
| Datasets | GET /datasets · POST /datasets/load-local · POST /datasets/upload · GET/DELETE /datasets/{id} |
| Images | GET /datasets/{id}/images · GET /datasets/{id}/images/{image_id} · PUT .../annotations |
| Classes | POST /datasets/{id}/extract-classes · /delete-classes · /merge-classes |
| Conversion | POST /datasets/{id}/convert · POST /datasets/merge · GET /formats |
| Augmentation | POST /datasets/{id}/augment-enhanced |
| Video | POST /video/extract |
| Duplicates | POST /datasets/{id}/find-duplicates · /remove-duplicates |
| Models | GET /models · POST /models/download · POST /models/import · POST /models/{id}/load · POST /models/{id}/unload |
| Auto-annotation | POST /datasets/{id}/auto-annotate · GET /api/auto-annotate/jobs |
| Training | POST /training/start · GET /training/{job_id}/status · /pause · /resume · /stop |
| System | GET /api/health |
"Backend not connected" — Python failed to start. Check .logs/backend.log. Common causes: Python < 3.10, port 8000 in use, missing OpenCV system dependency.
First startup hangs — Normal. PyTorch is large. Check .logs/backend.log to watch pip progress.
"Dataset format not recognized" — Auto-detection looks for specific files (data.yaml, instances_train.json, Annotations/*.xml, etc). Match the folder structure exactly. Nested ZIPs aren't supported — extract first.
OOM during training — Lower batch size (try 4–8), lower image size (try 320), or use a smaller arch (yolov8n). Check VRAM with nvidia-smi.
Port still in use after crash — python3 run.py stop. If that fails: lsof -ti:3000 | xargs kill -9 (macOS/Linux) or netstat -ano | findstr :3000 → taskkill /F /PID <pid> (Windows).
Blank frontend / 500 — Run npm install manually in the project root, then python3 run.py restart-front.
⚠️ The FastAPI backend serves files directly from your local filesystem. Don't expose port 8000 to the public internet without authentication. For remote GPU servers, use SSH port forwarding:ssh -L 3000:localhost:3000 -L 8000:localhost:8000 user@server