How woxihuanjiangguo/bevnext works
Project type and purpose
• BEVNeXt is a deep‐learning research framework for 3D object detection using Bird’s‐Eye‐View (BEV) representations.
• It is built on top of mmdetection3d and BEVDet, providing modular components for data handling, model definition, training, inference, and deployment.
• It’s neither a web service nor a classic full-stack app, but rather a machine‐learning library + CLI toolchain + model zoo.
File structure
• config/… : hundreds of Python config files grouped by model family (bevnext, bevdet, centerpoint, etc.), base defaults, datasets, schedules.
• mmdet3d/… : the core Python package with submodules
• apis: train/inference/test entry-points
• core: anchor generators, bbox coders, voxel utilities, hook system, evaluation
• datasets: dataset wrappers, pipelines, loading/augmentation
• models: backbones, necks, heads, detectors, losses, fusion layers, ops
• ops: custom CUDA/CPU kernels (bev_pool, spconv, dgcnn)
• utils: environment setup, logger, config compatibility
• tools/: command-line scripts for data conversion, training (dist_train.sh), testing (dist_test.sh), benchmarking, model conversion, deployment (TorchServe handler)
• demo/: example scripts for point-cloud demos
• docker/: Dockerfiles for building containers and serving models
• docs/: Sphinx documentation in English/Chinese
• tests/: unit tests covering datasets, pipelines, models, runtime, utils
README for additional insights
• Describes installation, dataset preparation, multi-stage training & inference workflows
• Highlights dependency on Anaconda, PyTorch, mmcv, mmdetection3d
• Points to config files in configs/bevnext/ for stage1 and stage2
Main components & their interactions
a. Data Layer
• Raw datasets (nuScenes, KITTI, Waymo…)
• tools/create_data_bevdet.py → preprocess → generate PKL files
• mmdet3d.datasets + pipelines: loading, augmentation, voxelization
b. Configuration Manager
• Python config files (configs/_base_, configs/bevnext/…)
• Registry & builder pattern (mmdet3d/models/builder.py, mmdet3d/datasets/builder.py)
c. Model Components
• Backbones (ResNet, PointNet2, DGCNN, Voxel encoders)
• Necks (FPN, scatter, view_transformer)
• Heads (dense heads, ROI heads, segmentation heads)
• Fusion modules (temporal fusion for BEV)
• Loss functions & post-processing (NMS, decode)
d. Training & Validation
• mmdet3d.apis.train → builds dataset, model, optimizer, hooks
• Hook system for EMA, sync-BN, checkpointing
• Distributed training scripts (tools/dist_train.sh)
e. Inference & Evaluation
• mmdet3d.apis.inference / test → load checkpoint → forward → postprocess → metrics
• tools/dist_test.sh, tools/deployment/mmdet3d2torchserve.py
f. Deployment
• Docker containers (docker/, docker/serve/)
• TorchServe handler (tools/deployment/mmdet3d_handler.py)
Architectural patterns & design principles
• Modular plugin registry (models, datasets, pipelines)
• Separation of concerns: data-I/O, model definition, training loop, deployment
• Config-driven instantiation: no hard-coded hyperparams
• Extendability: add new backbones/heads by registering in builder