oarriaga/paz

What it solves

PAZ is a hierarchical perception library for autonomous systems, designed to consolidate deep learning tools for robot perception using Keras/TensorFlow. It provides a structured way to build vision pipelines that range from simple out-of-the-box predictions to complex, custom-built perception systems.

How it works

PAZ organizes its functionality into three API levels to balance ease of use with flexibility:

• High-level API: Offers pre-implemented "pipelines" for immediate prediction (e.g., object detection using SSD512COCO).
• Mid-level API: Allows users to construct sequential or explicit pipelines using "Processors." Sequential processors can chain together functions like data augmentation, while explicit processors allow for custom logic (e.g., combining face detection and emotion classification).
• Low-level API: Provides the backend foundation with utility functions for handling boxes, cameras, images, keypoints, and quaternions.

Who it’s for

It is intended for developers and researchers working on autonomous systems and robotics, specifically those using the Python ecosystem (TensorFlow/Keras) who need a modular way to implement computer vision tasks.

Highlights

• Diverse Model Support: Includes implementations for object detection (SSD), semantic segmentation (U-NET), 6D pose estimation (Pix2Pose), and human/hand pose estimation (HRNet, DetNet).
• Modular Pipeline Architecture: Enables the creation of complex perception workflows by composing mid-level processors.
• Integrated Data Loaders: Built-in support for datasets such as OpenImages, VOC, CityScapes, and FERPlus.
• Robot-Ready: Includes built-in messages (e.g., Pose6D) to facilitate data exchange with frameworks like ROS.