An interactive full-stack dashboard for Segment Anything Model 3 (SAM3), providing a complete pipeline for image segmentation, model fine-tuning, data labeling, and performance evaluation on geospatial/satellite imagery.
- Year
- 2026
- Status
- In development
- Stack
- Next.js 16, React 19, TypeScript, TailwindCSS 4, DaisyUI 5, Fabric.js, Chart.js, FastAPI, SQLAlchemy, PyTorch, CUDA 12.6
What I built
- Built a complete SAM3 environment with Dashboard (system health), Inference (text/box/point prompts), Training (fine-tuning with TensorBoard), and Evaluation (IoU metrics).
- Implemented a 3-phase Data Engine pipeline: Batch Processing → Human Verification (accept/reject/flag) → Polygon Correction with Fabric.js editor.
- Developed a Polygon Editor with vertex editing, simplification, multi-select, keyboard shortcuts (V/E/H/D modes), undo/redo, and zoom/pan controls.
- Created Dataset management for uploading ZIP datasets (COCO format), converting image+mask folders, and validating local paths.
- Exports COCO-format datasets with RLE encoding ready for SAM3 training.
- Backend built with FastAPI, SQLAlchemy (SQLite), and OpenCV for image processing.
- Frontend built with Next.js 16, React 19, TypeScript, TailwindCSS v4, DaisyUI v5, Fabric.js, and Chart.js.
- Supports TIFF, PNG, JPG formats with real-time mask visualization overlays.