How facebookresearch/detectron works
Detectron is a Caffe2 research system for object detection and related instance tasks. Researchers configure a dataset, prepare image and region-of-interest data, build and train or run a detector, then evaluate or visualize predictions. The README names Mask R-CNN, RetinaNet, Faster R-CNN, RPN, Fast R-CNN and R-FCN, with ResNet/ResNeXt, FPN and VGG backbones. Source-index references support the internal connections shown; CLI wiring and several algorithm details are unsampled, so those links are marked unverified.
Main entry points: model_builder.py Model builder · infer.py Inference CLI · detector.py Detector runtime · task_evaluation.py Task evaluation
Components 25
Data and evaluation
- COCO dataset
detectron/datasets/json_dataset.py - Dataset catalog
detectron/datasets/dataset_catalog.py - Region database
detectron/datasets/roidb.py - Minibatch loader
detectron/roi_data/loader.py - Minibatch assembly
detectron/roi_data/minibatch.py - ROI targets
detectron/roi_data/fast_rcnn.py - Task evaluation
detectron/datasets/task_evaluation.py - COCO evaluation
detectron/datasets/json_dataset_evaluator.py - Cityscapes evaluation
detectron/datasets/cityscapes_json_dataset_evaluator.py - VOC evaluation
detectron/datasets/voc_dataset_evaluator.py
Detection models
- Model builder
detectron/modeling/model_builder.py - Detector runtime
detectron/modeling/detector.py - Backbones and FPN
detectron/modeling/FPN.py - ResNet backbone
detectron/modeling/ResNet.py - Anchor generation
detectron/modeling/generate_anchors.py - Detection heads
detectron/modeling - Proposal operations
detectron/ops/generate_proposals.py - FPN proposal routing
detectron/ops/collect_and_distribute_fpn_rpn_proposals.py
Training and inference
- Training CLI
tools/train_net.py - Inference CLI
tools/infer.py - Configuration
detectron/core/config.py
Core utilities
- Tensor and geometry utilities
detectron/utils
Other components
- Researcher
- Caffe2 framework
- Researcher results
Connections 28
- Researcher → Training CLI: starts training
- Researcher → Inference CLI: requests inference
- Training CLI → Model builder: builds model
- Inference CLI → Model builder: builds model
- Model builder → Configuration: reads settings
detectron/modeling/model_builder.py - COCO dataset → Dataset catalog: looks up metadata
detectron/datasets/json_dataset.py - COCO dataset → Region database: creates entries
detectron/datasets/roidb.py - Minibatch loader → Minibatch assembly: loads minibatches
detectron/roi_data/loader.py - Minibatch assembly → ROI targets: assembles targets
detectron/roi_data/minibatch.py - Model builder → Minibatch loader: uses loader
detectron/modeling/model_builder.py - Model builder → Minibatch assembly: uses minibatches
detectron/modeling/model_builder.py - Model builder → Detection heads: dispatches heads
detectron/modeling/model_builder.py - Model builder → Detector runtime: creates detector
detectron/modeling/model_builder.py - Model builder → Backbones and FPN: selects backbone
detectron/modeling/model_builder.py - Backbones and FPN → ResNet backbone: uses backbone
detectron/modeling/FPN.py - Backbones and FPN → Anchor generation: uses anchors
detectron/modeling/FPN.py - Detector runtime → ROI targets: uses ROI data
detectron/modeling/detector.py - Detector runtime → Proposal operations: generates proposals
detectron/modeling/detector.py - Detector runtime → FPN proposal routing: routes proposals
detectron/modeling/detector.py - FPN proposal routing → ROI targets: uses ROI targets
detectron/ops/collect_and_distribute_fpn_rpn_proposals.py - COCO dataset → Tensor and geometry utilities: uses geometry helpers
detectron/datasets/json_dataset.py - Training CLI → Caffe2 framework: runs on
- Inference CLI → Caffe2 framework: runs on
- Task evaluation → COCO evaluation: dispatches evaluator
detectron/datasets/task_evaluation.py - Task evaluation → Cityscapes evaluation: dispatches evaluator
detectron/datasets/task_evaluation.py - Task evaluation → VOC evaluation: dispatches evaluator
detectron/datasets/task_evaluation.py - Inference CLI → Task evaluation: evaluates predictions
- Inference CLI → Researcher results: returns detections