Pi3X EU AI Act Compliance Profile

$\pi^3$: Permutation-Equivariant Visual Geometry Learning

This repository contains the weights for Pi3X, an enhanced version of the $\pi^3$ model introduced in the paper $\pi^3$: Permutation-Equivariant Visual Geometry Learning.

$\pi^3$ is a feed-forward neural network for visual geometry reconstruction that eliminates the need for a fixed reference view. It employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps from an unordered set of images, making it robust to input ordering and achieving state-of-the-art performance.

• Project Page: yyfz.github.io/pi3/
• GitHub Repository: github.com/yyfz/Pi3
• Demo: Hugging Face Space

Pi3X Engineering Update

Pi3X is an enhanced version focusing on flexibility and reconstruction quality:

• Smoother Reconstruction: Uses a Convolutional Head to reduce grid-like artifacts.
• Flexible Conditioning: Supports optional injection of camera poses, intrinsics, and depth.
• Reliable Confidence: Predicts continuous quality levels for better noise filtering.
• Metric Scale: Supports approximate metric scale reconstruction.

Sample Usage

To use this model, you need to clone the official repository and install the dependencies.

import torch
from pi3.models.pi3x import Pi3X            # new version (Recommended)
from pi3.utils.basic import load_images_as_tensor 

# --- Setup ---
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Pi3X.from_pretrained("yyfz233/Pi3X").to(device).eval()

# --- Load Data ---
# Load a sequence of N images into a tensor (N, 3, H, W)
# pixel values in the range [0, 1]
imgs = load_images_as_tensor('path/to/your/data', interval=10).to(device)

# --- Inference ---
print("Running model inference...")

Your obligations as a Pi3X deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags