Faster R-CNN Object Detection on Open Images

This repository contains a PyTorch implementation of Faster R-CNN trained and evaluated on the Open Images Dataset. The project was originally developed in Google Colab and later refactored into a modular, reusable, and GitHub-friendly structure.

The code supports:

• Training Faster R-CNN on Open Images bounding box annotations
• Validation using mAP (Mean Average Precision)
• Inference and visualization on custom images
• Checkpoint saving and loading

---

Requirements

Create a Python environment (Python ≥ 3.9) and install dependencies. Main dependencies:

• torch
• torchvision
• torchmetrics
• pandas
• Pillow
• matplotlib
• tqdm

---

Dataset Preparation

This project uses the Open Images Dataset (bounding box annotations and image folders).

Required files:

• train-annotations-bbox.csv
• validation-annotations-bbox.csv
• class-descriptions-boxable.csv

Extracted image folders:

• train_x/ (with x = 1,..,f)
• validation/

Since the entire train dataset provided by Open Images is huge, it was split in several chunks (from train_0 to train_f).

---

Training

Run training and validation: python training/train.py

What happens:

• Loads Open Images annotations
• Trains Faster R-CNN (ResNet-50 FPN backbone)
• Evaluates on validation set using mAP
• Saves a checkpoint: fasterrcnn_checkpoint.pth

Checkpoint includes:

• Model weights
• Optimizer state
• Epoch number
• Validation metrics

---

Model Architecture

• Faster R-CNN
• Backbone: ResNet-50 + FPN
• Pretrained on COCO (optional)
• Custom classifier head for Open Images classes