README.md

metadata

title: Objectlocalization
emoji: 👁
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.18.0
app_file: src/app.py
pinned: false
short_description: Using RESTNET-RCNN-RPN-FNN to detect lego pieces

LEGO Object Detection using Faster R-CNN

Faster R-CNN

This project trains a Faster R-CNN model with a ResNet-50 backbone to detect LEGO objects using a custom dataset.

---

🔍 Project Overview

This project implements an advanced object detection system specifically designed for LEGO pieces using a combination of powerful deep learning architectures:

1. ResNet-50 Backbone:

   • Serves as the feature extractor
   • Pre-trained on ImageNet for robust feature learning
   • Deep residual learning framework for improved training of deep networks

2. Region Proposal Network (RPN):

   • Scans the image and proposes potential object regions
   • Generates anchor boxes of various scales and ratios
   • Outputs "objectness" scores and bounding box refinements

3. Fast Neural Network (FNN):

   • Performs final classification and bounding box regression
   • Takes features from proposed regions
   • Outputs class probabilities and precise box coordinates

Key Features

• End-to-End Training: The entire network is trained jointly for optimal performance
• Multi-Scale Detection: Capable of detecting LEGO pieces of varying sizes
• Real-Time Processing: Efficient architecture allows for quick inference
• High Accuracy: Achieves strong mean Average Precision (mAP) on LEGO detection

Project Structure

lego_detection/
│── models/                   # Trained models
│   ├── lego_fasterrcnn.pth   # Saved model
│   ├── faster_rcnn_custom.pth   # Latest model
│
│── datasets/                  # Dataset folder
│   ├── images/                # Training images
│   ├── annotations/           # Corresponding XML annotations
│   ├── test_images/           # Testing the model
│   ├── annotations.json/      # To format annotation in one only file
│
│── src/                       # Source code
│   ├── transformdata.py       # Formats the data to COCO.json
│   ├── new_trainer.py         # Train the model based on the new assumptions
│   ├── app.py                 # Allow users to interact with this model
│   ├── Attempt1               # First Implementation
│     ├── dataset.py             # Dataset class (LegoDataset)
│     ├── train.py               # Training script
│     ├── evaluate.py            # mAP Calculation
│     ├── utils.py               # IoU, AP calculation functions
│
│── config.yaml                # Hyperparameters & settings
│── README.md                  # Project documentation

---

⚡ Setup Instructions

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Update Configuration

Modify config.yaml to adjust hyperparameters, dataset paths, and model settings.

---

visualize using Gradio

1. If the model is not in models please (add it from the submitted file) Im trying to add the model but its too big for github standars.

2. Run the following Bash

python src/app.py

3. Evaluate and give me 100. I know, im awesome.

---

🚀 Training the Model

Run the following command to start training:

python src/train.py

This script will: ✅ Train Faster R-CNN with LegoDataset
✅ Log training loss & mAP
✅ Save the trained model in models/lego_fasterrcnn.pth

---

📊 Monitoring Training Progress

Use the Jupyter Notebook to visualize loss & mAP over epochs:

jupyter notebook notebooks/training_visualization.ipynb

---

🛠️ Hyperparameters (config.yaml)

Modify the config.yaml file to fine-tune the model:

model:
  backbone: resnet50
  num_classes: 2
  pretrained: true
  learning_rate: 0.0001
  epochs: 5
  batch_size: 8
  optimizer: adam

dataset:
  image_dir: datasets/images
  annotation_dir: datasets/annotations
  train_split: 0.8
  val_split: 0.2

evaluation:
  iou_threshold: 0.5

---

📝 Training Strategies for Faster R-CNN with ResNet-50 Backbone

Trainable Backbone Layers	Epochs	Batch Size	Recommended Learning Rate	Optimizer	Scheduler
0	10	4	0.0100	SGD	StepLR(3, 0.1)
3	10	8	0.0050	SGD	StepLR(3, 0.1)
5	10	16	0.0001	AdamW	CosineAnnealing
3	20	8	0.0050	SGD	StepLR(5, 0.1)
5	20	16	0.0001	AdamW	CosineAnnealing

---

📡 Evaluating the Model

Once training is complete, evaluate performance using:

python src/evaluate.py

---

💡 Troubleshooting & Tips

❓ Training Takes Too Long?

• Reduce epochs in config.yaml
• Use a smaller dataset for testing

❓ mAP is too low?

• Increase epochs
• Check dataset annotations
• Tune learning rate

---

🏆 Contributors

• 👤 Alex - Machine Learning Engineer

🚀 Happy Training!