kursatkomurcu/YOLO-MT-A-Lightweight-Multi-Task-Learning-Framework

YOLO-MT-A-Lightweight-Multi-Task-Learning-Framework-

You can find the inference and training code from bdd100k.ipynb notebook.

• —GitHub: https://github.com/kursatkomurcu/YOLO-MT-A-Lightweight-Multi-Task-Learning-Framework-

🚗 YOLO-MT: Lightweight Multi-Task Learning for Unified Scene Understanding

YOLO-MT is a YOLOv12-based lightweight multi-task learning framework that performs object detection, lane detection, drivable area segmentation, and scene attribute classification within a single unified model. It is specifically designed for real-time autonomous driving and embedded systems with limited computational resources.

✨ Key Features

• —✅ 4 tasks in a single network
• —Object Detection
• —Lane Detection
• —Drivable Area Segmentation
• —Scene Attribute Classification (weather, scene type, time of day)
• —⚡ Real-time inference
• —📦 Only 2.9M parameters
• —🧠 Shared YOLOv12 backbone
• —🔥 Optimized for embedded and edge devices
• —📊 Trained and evaluated on the BDD100K dataset

🏗️ Architecture Overview

• —Shared YOLOv12 encoder
• —Lightweight multi-branch decoder
• —ASPP-Lite module for context aggregation
• —Separate task-specific heads for:
• —Lane segmentation
• —Drivable area segmentation
• —Attribute classification

🎯 Training Strategy

1. 1.Train YOLOv12 on object detection
2. 2.Freeze the backbone
3. 3.Jointly train:
4. 4.Lane detection
5. 5.Drivable area segmentation
6. 6.Attribute classification

This two-stage training ensures strong feature reuse while keeping the model extremely compact.

📁 Dataset

• —BDD100K
• —Object detection annotations
• —Lane markings
• —Drivable area masks
• —Scene attributes (weather, scene type, time of day)

All images are resized to 384×640 for efficient real-time processing.

📌 Use Cases

• —Autonomous driving perception
• —Advanced Driver Assistance Systems (ADAS)
• —Embedded AI systems
• —Edge deployment for robotics

Example outputs