API Reference

This reference maps the documented method to the repository files that implement it. It is intentionally scoped to the active BEVFormerFusion path and excludes legacy or experimental files that are not treated as part of the public method.

Code-to-documentation mapping

Module	File path	Role in method	Description
Active config	projects/configs/bevformer/bevformer_project.py	Configuration root	Declares the published model, dataset wiring, loss configuration, and optimization schedule.
Detector	projects/mmdet3d_plugin/bevformer/detectors/bevformer.py	Train and test orchestration	Builds image and LiDAR features, manages temporal caches, and dispatches to the BEVFormer head.
Transformer	projects/mmdet3d_plugin/bevformer/modules/transformer.py	BEV construction and decoder fusion	Computes BEV tokens, projects LiDAR BEV maps, snapshots bev_embed_cam, and applies decoder-side fusion.
Encoder sequence	projects/mmdet3d_plugin/bevformer/modules/encoder.py	Multi-modal BEV refinement	Implements BEVFormerEncoder, BEVFormerLayer, and MM_BEVFormerLayer, including the dual-attention fusion path.
Camera cross-attention	projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py	Camera-to-BEV lifting	Projects BEV reference points into camera frames and performs image-side deformable attention.
Temporal attention	projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py	Temporal BEV update	Mixes current and previous BEV states through deformable attention over the BEV grid.
Decoder	projects/mmdet3d_plugin/bevformer/modules/decoder.py	Object-query decoding	Runs decoder self-attention and cross-attention over the fused BEV tensor.
Detection head	projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py	Multi-branch prediction	Produces class scores, boxes, yaw bins, yaw residuals, and dedicated velocity predictions.
BBox coder	projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py	Final decoding	Converts decoder outputs into final boxes and injects velocity-head predictions into the output boxes.
Dataset	projects/mmdet3d_plugin/datasets/nuscenes_dataset.py	Temporal sample assembly	Builds temporal queues, enriches img_metas, and forwards geometry metadata required by the transformer.

Public command surface

Entry point	Path	Purpose
Training CLI	tools/train.py	Launches model training from a config file.
Evaluation CLI	tools/test.py	Runs evaluation and optional BEV visualization.
Benchmark CLI	tools/analysis_tools/benchmark.py	Measures inference throughput when the runtime stack is available.
Dataset conversion	tools/create_data.py	Creates nuScenes temporal info files and related annotations.

Documentation roles

Documentation page	Primary code anchors
architecture.md	Detector, transformer, encoder, head, dataset
bevformer-comparison.md	Official BEVFormer config plus active config, transformer, encoder, head
experiments.md	Result workbooks, active config, canonical metrics JSON
usage.md	Active config, tools/train.py, tools/test.py, tools/create_data.py, tools/analysis_tools/benchmark.py

Legacy and excluded files

The following paths are intentionally excluded from the public method narrative:

Path	Reason
projects/configs/bevformerv2/	Separate experiment family, not the active BEVFormerFusion path.
projects/mmdet3d_plugin/bevformer/modules/transformer copy.py	Stale duplicate of the active transformer implementation.
projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head_old.py	Retired head variant.
tools/tests/test_petr3d_bevformer.py	Imports missing petr3d_embedding.py and therefore is not a valid regression test for the published method.

Interface notes

• No new public Python API is introduced by this documentation overhaul.
• The authoritative public outputs are the README, the MkDocs site rooted at /, and the canonical metrics artifact at docs/assets/data/metrics.json.
• The generated site/ directory remains a build artifact and is not the editable source of truth.