parkie0517/raf

How parkie0517/raf works

RAF is a research system for adverse-weather 3D object detection using camera, LiDAR, and 4D RADAR. The documented workflow trains a camera reliability branch and BEV fusion/detection components using pretrained LiDAR–RADAR networks, then evaluates detections. The graph preserves the dataset, model, training, and evaluation stages, plus core supporting utilities. Some model-stage links come from the README; source excerpts are partial, so those links are marked dashed where implementation wiring is not shown.

Main entry points: pipeline_detection_v1_1.py Train pipeline · pipeline_detection_v1_0.py Detection pipeline · crn_lidar.py LiDAR–RADAR backbone · util_pipeline.py Pipeline utilities

Components 22

Data and preparation

K-Radar v1 datasets/kradar_detection_v1_1.py
K-Radar v1.0 datasets/kradar_detection_v1_0.py
K-Radar v2.4 datasets/kradar_detection_v2_4.py
Point preparation models/pre_processor

Detection models

LiDAR–RADAR backbone models/skeletons/crn_lidar.py
Pillar encoding models/backbone_3d/vfe/pillar_vfe.py
BEV backbone models/backbone_2d/base_bev_backbone.py
Camera branch models/bevdepth/base_lss_fpn.py
BEV fusion models/fusers
Detection head models/head/center_head.py
3D NMS models/model_utils/model_nms_utils.py

Training and evaluation

Train pipeline pipelines/pipeline_detection_v1_1.py
Detection pipeline pipelines/pipeline_detection_v1_0.py
Detection metrics utils/kitti_eval/eval.py
Evaluation data utils/kitti_eval/kitti_common.py

Core support

Configuration utils/util_config.py
Pipeline utilities utils/util_pipeline.py
Calibration utils/util_calib.py
Geometry utilities utils/util_geometry.py
Optimization utils/util_optim.py
Point-cloud utilities utils/util_point_cloud.py

Other components

Researcher

Connections 27

Researcher → Train pipeline: trains model
Researcher → Detection pipeline: evaluates model
Train pipeline → Pipeline utilities: uses pipelines/pipeline_detection_v1_1.py
Train pipeline → Configuration: loads config pipelines/pipeline_detection_v1_1.py
Train pipeline → Optimization: uses pipelines/pipeline_detection_v1_1.py
Train pipeline → Point-cloud utilities: uses pipelines/pipeline_detection_v1_1.py
Train pipeline → Calibration: uses pipelines/pipeline_detection_v1_1.py
Train pipeline → K-Radar v1: loads data
Detection pipeline → Pipeline utilities: uses pipelines/pipeline_detection_v1_0.py
Detection pipeline → Configuration: loads config pipelines/pipeline_detection_v1_0.py
Detection pipeline → Point-cloud utilities: uses pipelines/pipeline_detection_v1_0.py
Detection pipeline → Calibration: uses pipelines/pipeline_detection_v1_0.py
Detection pipeline → Optimization: uses pipelines/pipeline_detection_v1_0.py
Detection pipeline → Detection metrics: evaluates with pipelines/pipeline_detection_v1_0.py
Detection metrics → Evaluation data: uses utils/kitti_eval/eval.py
K-Radar v1 → Geometry utilities: uses datasets/kradar_detection_v1_1.py
Pipeline utilities → K-Radar v1.0: references dataset utils/util_pipeline.py
Pipeline utilities → K-Radar v2.4: references dataset utils/util_pipeline.py
Pipeline utilities → LiDAR–RADAR backbone: builds model utils/util_pipeline.py
LiDAR–RADAR backbone → Point preparation: prepares points models/skeletons/crn_lidar.py
LiDAR–RADAR backbone → Pillar encoding: encodes pillars
LiDAR–RADAR backbone → BEV backbone: extracts BEV
LiDAR–RADAR backbone → Camera branch: adds camera branch
Camera branch → BEV fusion: fuses in BEV
LiDAR–RADAR backbone → BEV fusion: fuses modalities models/skeletons/crn_lidar.py
BEV fusion → Detection head: feeds detections
Detection head → 3D NMS: suppresses boxes models/head/center_head.py

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·