Object Detection Pipeline

This repository contains a modular pipeline for training object detection models. It currently supports Faster R-CNN and includes an implementation of DETR, allowing for easy switching between architectures and configurations.

Features

• Supported Models:
  • Faster R-CNN (ResNet, MobileNet, etc.)
  • DETR (Transformer-based detection)
• Configuration Management: All training parameters are controlled via centralized YAML files.
• Hyperparameter Optimization: Integrated support for Optuna.
• Data Format: COCO Style (JSON).

Directory Structure

The project is organized to separate the source code (src/) from the configuration (config/) and the execution scripts (located at the root).

.
├── config/             # YAML configuration files
├── data/               # Dataset (Images and labels.json)
├── src/                # Library source code (Models, Engine, Utils)
├── train.py            # Standard training script
├── train_optuna.py     # Hyperparameter optimization script, only for Faster-RCNN
├── eval.py             # Evaluation script (mAP metrics)
├── inference.py        # Inference script for new images
└── requirements.txt

Dataset Structure

The project expects data to be organized in a data/ folder. The labels.json file must follow the COCO Object Detection standard.

data/
├── train/
├── test/
├── validation/
└── labels.json

Below is an example of the expected JSON structure:

{
    "images": [
        {
            "id": 1,
            "file_name": "0566_1176902707_01_WRI-L2_M015.png",
            "width": 652,
            "height": 1072
        }
    ],
    "annotations": [
        {
            "id": 1,
            "image_id": 1,
            "category_id": 0,
            "bbox": [186, 468, 188, 77],
            "area": 14476,
            "iscrowd": 0
        }
    ],
    "categories": [
        {
            "id": 0,
            "name": "object"
        }
    ]
}

Installation

Install the required dependencies:

pip install -r requirements.txt

Usage & Execution

All execution scripts are located at the root of the repository.

1. Standard Training (train.py)

The main entry point for training. It loads the configuration, initializes the model and dataset, and runs the training loop with logging and checkpointing.

# Train Faster R-CNN
python train.py fasterrcnn 

# Train DETR
python train.py detr 

2. Hyperparameter Optimization (train_optuna.py)

A dedicated script that uses Optuna to find the best hyperparameters (e.g., learning rate, weight decay). It runs multiple trials based on the search space defined in the config.

python train_optuna.py

3. Evaluation (eval.py)

Evaluates a trained model against the test set. It calculates standard metrics such as mAP (Mean Average Precision) following COCO standards.

python eval.py fasterrcnn(or DETR)

4. Inference (inference.py)

Used to run object detection on new, unseen images. It generates visual results with bounding boxes drawn over the detected objects.

python inference.py fasterrcnn(or DETR)

Configuration (YAML)

Modify the files in config/ to adjust your experiments:

• config_faster.yaml: Standard configuration for Faster R-CNN.
• config_detr.yaml: Base configuration for DETR.