pax-detector 0.1.1

A vision pipeline for object and car make classification with CLI

Pax Case: Object and Car Make Classification

This project provides a vision pipeline for object and car make classification. It includes a command-line interface (CLI) for processing images and a Gradio-based web interface for interactive demonstrations.

Problem Statement

The goal is to build a tool that can:

1. Object Classification: Identify whether an image contains a car, truck, bicycle, or person.
2. Car Make Classification: If the image contains a car, identify its make (e.g., Volkswagen, Chevrolet).

The solution must be scalable to handle up to 10 million images daily and be flexible enough to accommodate future classification tasks.

Features

• Two-Stage Pipeline:
  • Object Detection: Uses a YOLOv8 model to detect objects (person, bicycle, car, truck).
  • Car Make Classification: Uses a Google SigLIP model fine-tuned on the Stanford Cars dataset to classify the make of detected cars.
• CLI Tool (detector):
  • Process single images or directories of images from the command line.
  • Supports configuration via a YAML file.
  • Adjustable confidence threshold for object detection.
• Scalability and Extensibility:
  • The architecture is designed to be scalable and extensible. See SCALING_STRATEGY.md for a detailed plan on scaling to 10M images/day and adding new classification types.

Package usage

1. Install PyPi package

pip install pax-detector

Local Setup and Installation

1. Clone the repository:

   git clone <repository-url>
   cd pax-case
   

2. Install dependencies: It is recommended to use a virtual environment. This project uses uv https://docs.astral.sh/uv/getting-started/installation/ for package management.

   uv venv 
   source .venv/bin/activate
   
   uv pip install -e . # To build package locally
   uv sync # If want to install local dependencies 
   

Usage

Command-Line Interface (CLI)

The CLI tool (detector) is used for processing individual images.

Basic Usage:

To process a single image:

detector --image_path /path/to/your/image.jpg

To process all images in a directory:

detector --images_dir /path/to/your/images/

Using a Configuration File:

You can also run the CLI using a config.yml file to specify parameters.

Example config.yml:

image_path: '/path/to/your/image.jpg'
confidence: 0.6

Run with config:

detector --config config.yml

Arguments:

• --image_path: Path to a single input image.
• --images_dir: Path to a directory of images.
• --config: Path to a YAML configuration file.
• --confidence: Confidence threshold for object detection (default: 0.5).

Example:

detector --image_path datasets/car-camera/images/00001.jpg --confidence 0.4

This will process the image and print the JSON output to the console.

Extensibility

For details on how to add more classification types (e.g., helmet color, t-shirt color), please refer to the SCALING_STRATEGY.md document, which outlines the proposed architecture for extending the pipeline.