How yujeong-star/rl_3dod works
Project type and purpose
• This is a research-grade 3D object‐detection codebase (“RL_3DOD”) integrating LiDAR and 4D radar data.
• It provides end‐to‐end data ingestion, preprocessing, model definition (multiple backbones, heads), training stages (image classification pretrain + detection), inference, and evaluation.
File structure
• Top‐level:
• README.md & setup.py – project description and install entry point
• configs/ – YAML and Python config templates (general, UI, radar fusion)
• datasets/ – dataset loader for K-Radar detection
• pipelines/ – high‐level detection pipeline script
• models/ – modular network definitions:
– backbone_2d (image or radar BEV), backbone_3d (sparse conv & fusion), heads, preprocessors, ROI heads, skeletons
– img_cls/ for two‐stage pretraining (cls_train.py, cls_test.py)
• resources/ – raw radar/LiDAR data mats & train/test splits
• main_train_0.py & main_cond_0.py – entry points for detection training and conditional testing
• Notice modular pattern: each concern (data, model, training, evaluation) is in its own directory. Utilities are further subdivided by functionality.
README for additional insights
• Two‐stage training: Stage1 image classification pretrain, Stage2 detection training
• Single‐script inference (main_cond_0.py)
• Relies on PyTorch, spconv, Open3D, OpenCV, custom CUDA Ops (Rotated_IoU)
• Based on K-Radar dataset
Components
a. Main components:
1. Data Sources:
• LiDAR point clouds (resources/*.mat)
• 4D Radar data (resources/*.mat)
• Train/Test splits (resources/split/)
2. Data Loader & Pre‐processor:
• datasets/kradar_detection_v1_0.py
• models/pre_processor/rdr_sparse_processor*.py
3. Configuration Manager:
• configs/*.yml, configs/*.py
4. Model Zoo:
• Backbone2D, Backbone3D (rl_3df.py), Fusion skeletons
• ROI & detection heads (rdr_spcube_head.py)
• Classification branch (img_cls/cls_model.py)
5. Training Pipelines:
• Stage1: img_cls/cls_train.py → produce pretrained weights
• Stage2: main_train_0.py & pipelines/pipeline_detection_v1_0.py
6. Inference & Evaluation:
• main_cond_0.py
• utils/kitti_eval/ for metrics
7. Utilities (shared):
• utils/Rotated_IoU CUDA ops
• utils/box_*, util_geometry, util_pipeline, util_point_cloud, UI vis
b. Relationships & data flows:
• Configs feed Data Loader & Model definitions
• Data Loader + Pre‐processor → Batch Tensors → Model
• Pretrained weights from Stage1 injected into Detection model in Stage2
• During inference, model outputs → post‐processing (NMS, RotatedIoU) → Evaluation
c. Architectural patterns and design principles:
• Modularity: each layer of the ML pipeline is its own package
• Two‐stage training flow (pretrain → fine‐tune)
• Separation of concerns: data, model, training, evaluation, utilities
d. Key technologies, frameworks, libraries:
• Python 3.8, PyTorch 1.10.1, spconv-cu111
• Open3D, OpenCV, numba
• Custom CUDA (rotate IoU)
• KittiEval scripts for metrics