Vision pipelines for the impatient and the tinkerer

What it does

PAZ is a Python perception library built around three API layers. At the top, ready-made pipelines for object detection, face recognition, pose estimation, and semantic segmentation accept a NumPy array and return predictions. The mid-level provides composable Processor and SequentialProcessor blocks for wiring custom augmentation or inference chains. At the bottom, a backend of raw functions handles boxes, keypoints, quaternions, and camera math. The entire library sits on only TensorFlow 2, OpenCV, and NumPy.

The interesting bit

The library treats every perception task as a pipeline of discrete blocks, so a high-level emotion detector is literally a subclassed Processor chaining a Haar cascade, a cropper, and a classifier. You can peel the abstraction layers like an onion instead of discovering hidden magic.

Key highlights

• High-level pipelines span a broad research spectrum: 6D pose estimation, probabilistic keypoints, Mask-RCNN, prototypical networks, MAML, and Eigenfaces.
• Mid-level SequentialProcessor chains accept arbitrary NumPy or OpenCV functions alongside library-native blocks.
• Built-in message types like Pose6D are designed to slot into ROS workflows without translation layers.
• Most included models can be re-trained on custom data, though the README flags Mask-RCNN re-training as in-progress.
• Dataset loaders are provided for VOC, OpenImages, FER2013, CityScapes, YCB-Video, and FAT.

Caveats

• Mask-RCNN re-training is explicitly flagged as in-progress; the example pipeline exists but custom training is not yet supported.

Verdict

PAZ suits robotics or CV researchers who want to prototype with pre-trained models and then surgically replace individual pipeline stages without leaving the same import tree. If you need a tightly scoped, production-hardened framework with a single abstraction level, this breadth might feel like overkill.