How maudzung/yolo3d-yolov4-pytorch works
Project Type and Purpose
• This is a deep-learning application for real-time 3D object detection on LiDAR point clouds, based on the YOLO3D (ECCV 2018) algorithm enhanced with YOLOv4 features.
File-Structure Overview
• dataset/
• Stores KITTI data (point clouds, labels, calibration, splits)
• docs/
• src/
• config/ – all hyperparameter and model‐CFG definitions (YOLOv4 cfg, train and dataset configs)
• data_process/ – conversion of raw LiDAR to BEV, data augmentation, train/val split, DataLoader
• models/ – Darknet→PyTorch conversion, model‐building utilities, YOLO layers
• utils/ – logging, evaluation-metric functions (IoU, GIoU), visualization, distributed training helpers
• train.py, test.py, evaluate.py (+ .sh wrappers) – entrypoints for training, inference, evaluation
• requirements.txt, README.md, LICENSE, .gitignore
1. Data Storage
• Raw KITTI Velodyne, labels, calib, images
• Train/val/test split files
2. Data Processing Pipeline
• BEV Encoder (height, intensity, density channels)
• Augmentations (Mosaic, Cutout, random scale/rotation, flip)
• Dataset & DataLoader (PyTorch)
3. Model Definition
• Backbone: CSPDarknet with Mish activations, SPP, PAN (YOLOv4)
4. Training Engine
• Loss calculation (MSE or GIoU), optimizer (SGD/Adam), scheduler
• Distributed DataParallel (NCCL)
• Checkpoint save/load
• Tensorboard logging
5. Inference & Evaluation
• Pretrained checkpoint loader
• Forward pass, NMS, thresholding
• Metric computation (AP, IoU)
6. Visualization
• Tensorboard dashboards
1. Data Storage → Data Processing
2. Data Processing → DataLoader → Training Engine
3. Training Engine ↔ Model Definition
4. Training Engine → Checkpoints
5. Checkpoints → Inference/Evaluation
6. Inference/Evaluation → Visualization & Metrics
7. Training Engine → Tensorboard
C. Architectural Patterns & Principles
• Modular separation: config / data_process / models / utils
• Config-driven: .cfg files and Python config modules decouple hyperparameters
• Reusable utilities: common IoU, logging, visualization functions
• DDP pattern for multi-GPU/multi-node training
D. Key Technologies & Frameworks
• Python 3.6+, PyTorch 1.5 (DDP via NCCL)
• NumPy, OpenCV (BEV processing), Mayavi, Shapely (visualization)
• Tensorboard for logging
• Shell scripts (train.sh, test.sh) for reproducibility
Detail Level
• Module (.cfg files, augmentations, YOLO layers, loss functions).
• How configurations branch into both data preprocessing and model instantiation.
• Don’t omit auxiliary utils (logger, misc, torch_utils) – they underpin the engine.