mrnabati/centerfusion

How mrnabati/centerfusion works

CenterFusion is a radar-and-camera 3D object detection system built around center-point image detections and frustum-based radar association. Its main workflow is dataset selection and preparation, model training or inference, then decoding and post-processing detections for evaluation or visualization. The graph also preserves documented nuScenes conversion, tracking, and dataset support. Source excerpts show key runtime calls; some subsystem links are documented by the README or source-index references rather than fully sampled implementation.

Main entry points: main.py Train / evaluate · detector.py Inference detector · generic_dataset.py Sample preparation · demo.py Demo entry

Components 24

Runtime workflow

Train / evaluate src/main.py
Run options src/lib/opts.py
Demo entry src/demo.py
Inference detector src/lib/detector.py
Experiment logging src/lib/logger.py

Data and datasets

Dataset registry src/lib/dataset/dataset_factory.py
Dataset adapters src/lib/dataset/datasets
Sample preparation src/lib/dataset/generic_dataset.py
nuScenes conversion src/tools/convert_nuScenes.py
Radar processing src/lib/utils/pointcloud.py
Image transforms src/lib/utils/image.py

Detection model

Model construction src/lib/model/model.py
Network architectures src/lib/model/networks
Fusion heads src/lib/model/networks/base_model.py
Detection decoding src/lib/model/decode.py
Training loop src/lib/trainer.py
Detection losses src/lib/model/losses.py

Results and evaluation

3D post-processing src/lib/utils/post_process.py
3D geometry src/lib/utils/ddd_utils.py
Object tracking src/lib/utils/tracker.py
Detection visualization src/lib/utils/debugger.py
Tracking evaluation src/tools/eval_kitti_track/evaluate_tracking.py

Other components

Researcher
Dataset files

Connections 27

Researcher → Train / evaluate: configures run README.md
Researcher → Demo entry: starts demo src/demo.py
Train / evaluate → Run options: parses options src/main.py
Train / evaluate → Dataset registry: selects dataset src/main.py
Train / evaluate → Model construction: creates model src/main.py
Train / evaluate → Training loop: runs training src/main.py
Train / evaluate → Experiment logging: logs metrics src/main.py
Demo entry → Run options: parses options src/demo.py
Demo entry → Inference detector: invokes inference src/demo.py
Dataset registry → Dataset adapters: selects adapter src/lib/dataset/dataset_factory.py
Dataset adapters → Sample preparation: inherits preparation src/lib/dataset/datasets/coco.py
Dataset files → Sample preparation: provides samples README.md
nuScenes conversion → Dataset files: converts annotations src/tools/convert_nuScenes.py
Sample preparation → Image transforms: transforms images src/lib/dataset/generic_dataset.py
Sample preparation → Radar processing: loads radar data src/lib/dataset/generic_dataset.py
Train / evaluate → Run options: sets dataset heads src/main.py
Model construction → Network architectures: builds architecture src/lib/model/model.py
Network architectures → Fusion heads: extends base src/lib/model/networks/dla.py
Training loop → Detection losses: computes losses src/lib/trainer.py
Training loop → Detection decoding: decodes outputs src/lib/trainer.py
Inference detector → Model construction: loads model src/lib/detector.py
Inference detector → Detection decoding: decodes predictions src/lib/detector.py
Inference detector → 3D post-processing: post-processes detections src/lib/detector.py
3D post-processing → 3D geometry: uses 3D geometry src/lib/utils/post_process.py
Inference detector → Object tracking: updates tracks src/lib/detector.py
Inference detector → Detection visualization: renders detections src/lib/detector.py
Train / evaluate → Tracking evaluation: evaluates tracking

More Python diagrams: rhoban/onshape-to-robot · zh-plus/openlrc · anssi-fr/bmc-tools · assetnote/php2shell-scanner · pipelex/pipelex

Drawn by GitDiagram on September 30, 2026 from the repository's file tree, README and source files. Read as Markdown ·