YOLOv8 Object Detection - Old Building Objects

Custom object detection model trained on YOLOv8 to identify objects in old building environments.

Project Overview

This project trains a YOLOv8 model to detect 4 classes of objects commonly found in old building environments:

• Door
• Dustbin
• Outdoor Plant
• Window

Technologies

• Python 3.12
• PyTorch 2.9.0
• Ultralytics YOLOv8
• CUDA 12.6
• Google Colab

Dataset

• Source: Roboflow - LGU Old Building Objects
• Classes: 4
• Format: YOLOv8
• Images: 760+ annotated images
• Train/Val/Test Split: 70/20/10

Model Architecture

• Base Model: YOLOv8n (nano) - lightweight and fast
• Input Size: 640x640
• Epochs: 50
• Device: NVIDIA Tesla T4 (Google Colab)

Training Results

Metric	Value
Precision (Door)	0.927
Recall (Door)	0.75
mAP50 (Door)	0.748
Precision (Dustbin)	0.893
Recall (Dustbin)	0.909
mAP50 (Dustbin)	0.906
Precision (Outdoor Plant)	0.893
Recall (Outdoor Plant)	0.909
mAP50 (Outdoor Plant)	0.906

Inference

The model successfully processes video input in real-time:

• Input: original.mp4 (1722 frames)
• Output: Annotated video with bounding boxes
• Speed: ~10ms per frame (100+ FPS on T4)
• Confidence Threshold: 0.25

Files

File	Description
src/train.py	Training script
src/predict.py	Inference script
models/best.pt	Trained model weights
data/data.yaml	Dataset configuration
videos/demo_output.mp4	Sample detection output

Author

Mahmoona Khan

Installation

pip install -r requirements.txt