ljnfwea/Aegis_Debris-Removal

Object Detection System

A comprehensive object detection system built with TensorFlow/Keras supporting custom model training and inference on images, videos, and real-time webcam feeds.

Features

• Multiple Backbones: MobileNetV2, ResNet50, EfficientNet
• Custom Training: Train on your own dataset
• Flexible Inference: Images, videos, and real-time webcam detection
• Easy Configuration: YAML-based configuration management
• Production Ready: Modular design with proper error handling

Project Structure

The-void/
├── main.py              # Main entry point
├── requirements.txt     # Python dependencies
├── config.yaml         # Configuration file
├── README.md           # This file
├── src/                # Source code
│   ├── config.py       # Configuration management
│   ├── dataset.py      # Dataset loading and preprocessing
│   ├── model.py        # Model definition
│   ├── train.py        # Training script
│   ├── detect.py       # Inference module
│   ├── evaluate.py     # Evaluation metrics
│   └── utils.py        # Utility functions
├── data/               # Training data
│   ├── train/         # Training images and annotations
│   ├── val/           # Validation images and annotations
│   └── test/          # Test images
├── models/            # Trained models
└── outputs/           # Detection results and logs

Installation

Prerequisites

• Python 3.8+
• CUDA 11.0+ (optional, for GPU acceleration)

Setup

1. Install dependencies:

pip install -r requirements.txt

Configuration

Edit config.yaml to customize training and inference parameters:

# Model settings
model:
  backbone: "mobilenetv2"  # Options: mobilenetv2, resnet50, efficientnet
  input_shape: [416, 416, 3]
  num_classes: 80

# Training settings
training:
  epochs: 100
  batch_size: 32
  learning_rate: 0.001

# Inference settings
inference:
  confidence_threshold: 0.5
  iou_threshold: 0.4

Quick Start

Training

Train with dummy data (for testing):

python main.py train --dummy-data

Train with custom data:

python main.py train --config config.yaml

Inference

Detect objects in an image:

python main.py detect --model models/final_model.h5 --input-image test.jpg

Detect objects in a video:

python main.py detect --model models/final_model.h5 --input-video test.mp4 --output-video output.mp4

Real-time detection from webcam:

python main.py detect --model models/final_model.h5 --webcam --duration 30

Dataset Format

The system supports Pascal VOC XML format for annotations:

<?xml version="1.0"?>
<annotation>
  <object>
    <name>person</name>
    <bndbox>
      <xmin>100</xmin>
      <ymin>150</ymin>
      <xmax>300</xmax>
      <ymax>450</ymax>
    </bndbox>
  </object>
</annotation>

Model Architecture

• Backbone: Transfer learning from pre-trained models (MobileNetV2, ResNet50, EfficientNet)
• Detection Head: Custom layers for bounding box regression and classification
• Outputs:
  • Bounding boxes: [x1, y1, x2, y2]
  • Confidence score: Object presence probability
  • Class predictions: 80-class softmax output

Evaluation Metrics

• IoU (Intersection over Union): Box overlap measurement
• Precision-Recall Curve: True positive and false positive trade-off
• Average Precision (AP): Area under precision-recall curve
• mAP: Mean average precision across all classes

Troubleshooting

Out of Memory Error

• Reduce batch_size in config.yaml
• Use smaller backbone (MobileNetV2)
• Reduce input image size

Low Accuracy

• Increase training data
• Train for more epochs
• Adjust detection thresholds
• Enable data augmentation

Advanced Usage

Custom Training

from src.train import train_model
from src.config import Config

config = Config('config.yaml')
model, history = train_model(config, use_custom_data=True)

Custom Dataset Loading

from src.dataset import ObjectDetectionDataset

dataset = ObjectDetectionDataset(
    image_dir='data/train/images',
    annotation_dir='data/train/annotations',
    image_size=(416, 416),
    class_names=['person', 'car', 'dog']
)

train_gen = dataset.get_data_generator(batch_size=32, shuffle=True)

Dependencies

• TensorFlow 2.15+
• NumPy
• OpenCV
• Pillow
• Matplotlib
• scikit-learn