apple/ml-hypersim resource

Health v2 · maintenance only

Flags: no_releases no_license

Adoption not part of the score

2043 stars · 153 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Hypersim is a photorealistic synthetic dataset of 74,619 images across 461 indoor scenes with dense per-pixel semantic instance segmentation labels and ground truth geometry. It also includes a toolkit for generating such images from publicly available 3D assets.

Use cases

• train semantic segmentation models for indoor scenes
• get per-pixel ground truth labels unavailable from real images
• train depth estimation or surface normal prediction models
• research intrinsic image decomposition into reflectance and illumination
• benchmark holistic indoor scene understanding models
• generate custom synthetic training images from 3D assets

When to choose

• you need dense per-pixel ground truth labels for indoor scene understanding
• you want complete scene geometry, material, and lighting information
• you need large-scale photorealistic training data with camera parameters
• you want to study view-dependent lighting effects via decomposed image components

When to avoid

• you need real-world photographic data rather than synthetic renders
• you cannot store roughly 1.9TB of image data
• your task involves outdoor scenes or people (people images were excluded)
• you need a small, lightweight dataset for quick experiments

Facets

dataset · maturity stable

data-generation image-processing computer-vision machine-learning computer-vision image-processing machine-learning artificial-intelligence simulation python windows synthetic-data scene-understanding semantic-segmentation depth-estimation indoor-scenes photorealistic-rendering per-pixel-labels 3d-assets linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="apple/ml-hypersim")