zakskyfighter/RGB_BEV_Kitti_Custom_Yolov12n

YOLOv12n LiDAR BEV Object Detection Model

Model Overview

This is a custom-trained YOLOv12n model for object detection on Bird’s Eye View (BEV) RGB images generated from LiDAR 3D point cloud data. The dataset used for training is derived from the KITTI dataset, converted from raw LiDAR point cloud data to 2D BEV images.

Dataset

• Source: KITTI Dataset
• Preprocessing: LiDAR point clouds converted into 2D RGB BEV images
• Custom Labels: Created for training

Training Details

• Training Platform: Kaggle Notebook
• Epochs: 300 (Continual learning)
• Batch Size: 32
• Input Image Size: 608 × 608
• Compute: 2× NVIDIA T4 GPUs (Distributed Training)
• Training Time: 14.5 hours
• Optimizer: AdamW

Data Augmentation & Training Arguments

The model was trained with the following augmentations and hyperparameters:

results = model.train(
    data=os.path.join(Dataset_folder, "data.yaml"),
    epochs=500,
    imgsz=608,
    plots=True,
    batch=batch_size,
    save=True,
    save_period=100,
    device="cuda",
    workers=4,
    project=Folder_name,
    seed=2005,
    copy_paste=0.15,
    optimizer="AdamW",
    mosaic=1.0,
    scale=0.9,
    verbose=True,
    resume=True,
    patience=100,
    cache=True,
    amp=True
)

Usage

To use this model for inference, load it using the Ultralytics YOLOv12 framework:

from ultralytics import YOLO

model = YOLO("path/to/your/yolov12n.pt")
results = model("path/to/your/image.jpg")
results.show()

Performance & Applications

• Designed for autonomous driving and LiDAR-based perception
• Capable of detecting objects from BEV RGB images derived from 3D LiDAR data
• Suitable for real-time object detection in self-driving applications

License

• mit

language

• english

metrics

• mean_iou

pipeline_tag

• object-detection

tags

• autonomous
• selfdriving
• LiDaR
• Kitti