FTC Vision

FTC Vision is an object detection project designed for the 2024-2025 FIRST Tech Challenge (FTC) season. This repository provides both PyTorch and TensorFlow implementations, enabling flexible training, validation, and inference workflows.

Key Features

• Multi-Framework Support: Implementations in both PyTorch and TensorFlow for training and inference.
• Dataset: Includes annotations and images of FTC game pieces. Available in VOC format and as TensorFlow-ready TFRecord files.
• TFLite Export: TensorFlow models can be exported to TFLite for deployment on lightweight devices.
• Comprehensive Tools: Utilities for preprocessing, dataset generation, and model conversion between frameworks.

Resources

The Resources used in this project is hosted on Hugging Face and is accessible at the links below:

Resources	Description
FTC Vision	Annotated dataset in VOC format, split into train/val with subdirectories for each class. Includes train/val TFRecord files and a label map.
FTC Vision - PyTorch	PyTorch implementation of the FTC Vision model, including training scripts and model weights.
Training Docs	Complete documentation for training of the PyTorch implimentaiton of FTC Vision

Repository Structure

.
├── DOCS/                         # Repository documentation
├── src_pytorch/                  # PyTorch implementation
├── src_tf/                       # TensorFlow implementation
├── utils/                        # Utility scripts for model training and evaluation
├── README.md                     # Project overview
└── requirements.txt              # Dependencies for the project

Start Here

Start by setting up your development environment:

Environment Setup

Demo Notebook

PyTorch Model Archetecture