Skip to content

ROADMAP.md

Latest commit

 

History

History
65 lines (50 loc) · 2.74 KB

ROADMAP.md

File metadata and controls

65 lines (50 loc) · 2.74 KB
Raw

SynthDet Roadmap

Completed Phases

Phase 1: Foundation + Analysis

  • Core types (BBox, ImageRecord, Dataset, SynthesisStrategy)
  • YOLO dataset loader, statistics, strategy generation
  • CLI: python -m synthdet.analysis

Phase 2: Compositor Pipeline

  • Defect patch extraction, clean background generation, Poisson blending
  • Bbox-aware classical augmentation (albumentations)
  • YOLO label writer
  • CLI: python -m synthdet.generate

Phase 3: API-Based Inpainting (Google Imagen 3)

  • Mask placement, inpainting pipeline, Imagen 3 provider
  • Local Stable Diffusion alternative (zero-cost)
  • Rate limiting, cost estimation, dry-run mode

Phase 3.5: Alternative Generation Pipelines

  • Generative Compositor: API-generated isolated patches + Poisson blending
  • Modify-and-Annotate: Whole-image transformation + auto-annotation

Phase 4: Auto-Annotation + Verification

  • Grounding DINO and OWL-ViT zero-shot detection
  • SAM-based bbox refinement
  • CLIP-based annotation quality scoring

Phase 5: Pipeline Orchestration

  • End-to-end pipeline with config-driven method selection
  • Dataset validation (bbox sanity, image integrity, split balance)
  • CLI: python -m synthdet.pipeline

Phase 6a: SPC Quality Monitoring + Embedding Diversity

  • Shewhart X-bar control charts with Western Electric rules
  • PyTorch forward-hook activation capture
  • CLIP/DINOv2 embedding-based diversity analysis

Phase 6b: Active Learning Loop

  • YOLO trainer wrapper, per-bucket/region evaluation
  • N-iteration generate → train → evaluate → refine coordinator

Phase 6c: Copy-Paste Augmentation + Web Acquisition

  • Copy-paste augmentation: Paste defect patches onto existing annotated images with Poisson blending (complements compositor which uses clean backgrounds)
  • Web scraper: Acquire background images from Google/Bing via icrawler with resolution filtering and perceptual deduplication
  • Relevance filter: CLIP-based filtering to keep only domain-relevant acquired images

Remaining Work

Mosaic / MixUp Augmentation

  • augment/mosaic.py: 4-image mosaic and MixUp for detection
  • Standard YOLO augmentation technique for small-object performance

Style Transfer

  • augment/style.py: Neural style transfer for domain adaptation
  • Transfer visual styles between datasets to improve generalization

Future Directions

  • Multi-class expansion: Extend beyond scratches to stains, broken bezels, dents
  • Dataset scaling: Larger base datasets with more diverse laptop models
  • Model zoo: Pre-trained weights for common defect detection tasks
  • Evaluation dashboard: Rich visualization of training progress, data quality, and model performance
  • Edge deployment: Model optimization (quantization, pruning) for real-time grading