Skip to main content

> ML_LIBRARY // DETECTRON2_v1.0

Detectron2

Meta FAIR (Fundamental AI Research) — Meta FAIR's modular deep learning platform for object detection and instance segmentation.

computer-visionv0.6Apache-2.0qualified
Official DocumentationSource Repository

Model Training

Supported
Accelerators:
CPUCUDA
Distributed Training:Yes

Model Inference

Supported
Inference Accelerators:
CPUCUDA
Deployment Targets:server

What It Does

  • +High-accuracy two-stage object detection (Faster R-CNN, Cascade R-CNN)
  • +State-of-the-art instance and panoptic segmentation (Mask R-CNN, PointRend)
  • +DensePose 3D surface mapping on human bodies
  • +Modular configuration system using yacs and fvcore

What It Does Not Do

  • -Natively export to ONNX without custom C++ op registration
  • -Run at real-time speeds on battery-powered mobile edge processors
  • -Receive frequent active feature updates (largely in maintenance mode)

>Suitable Work Types

  • High-precision document layout analysis (detecting tables, headers, signatures)
  • Medical tissue and pathology segmentation
  • Academic computer vision research benchmarks

>Unsuitable Work Types

  • Sub-10ms real-time edge video streaming (use YOLO or TensorRT)
  • Web browser client-side execution
Data Residency Implications

Operates entirely locally in memory. Zero external calls.

Security Considerations

Ensure build environment uses trusted CUDA compilers and dependencies.

Operational Profile & Known Limitations

Maturity:mature
Learning Curve:high
Ops Complexity:high
Cost Tier:free-oss
> Known Limitations:
  • Maintenance velocity has slowed in recent years; installation frequently conflicts with modern PyTorch/CUDA builds due to C++ extensions.

Associated Incident Patterns (Incidentpedia)

Enforce safeguards and monitoring to guard against these documented real-world failure modes:

cuda-driver-torch-mismatch

> Primary Evidence & Benchmark Citations

Detectron2 Documentationofficial-docs • 0.6
2026-09-25HIGH