hugsam/Room_Envelopes

Room Envelopes

See project page for details on this repository https://sambahrami.com/room_envelopes

Original dataset https://github.com/apple/ml-hypersim License https://creativecommons.org/licenses/by-nc-sa/3.0/

Modifications include creating point clouds from the Hypersim dataset and re-rendering depth maps with all objects removed, retaining the first structural layout surface for each example and its normal map.

This dataset is derived from the Hypersim dataset by Apple Inc., licensed under CC BY-NC-SA 3.0.

Usage

from datasets import load_dataset

# Load a specific split
ds = load_dataset("hugsam/Room_Envelopes", split="train")

# Or load all splits at once
ds = load_dataset("hugsam/Room_Envelopes")
# ds["train"], ds["validation"], ds["test"]

# Streaming (recommended for large datasets)
ds = load_dataset("hugsam/Room_Envelopes", split="train", streaming=True)

# Each sample is a dict:
sample = next(iter(ds))

sample["image"]          # PIL Image — original rendered RGBA image (1024×768)
sample["depth"]          # PIL Image — first-surface depth map (16-bit grayscale)
sample["normal"]         # PIL Image — first-surface normal map (16-bit RGB)
sample["layout_depth"]   # PIL Image — layout depth map (16-bit grayscale)
sample["layout_normal"]  # PIL Image — layout-layer normal map (16-bit RGB)
sample["seen_mask"]      # PIL Image — visibility mask (8-bit grayscale)
sample["json"]           # dict with metadata:
#   sample["json"]["scene"]       — scene identifier, e.g. "ai_001_001"
#   sample["json"]["camera"]      — camera identifier, e.g. "cam_00"
#   sample["json"]["frame"]       — frame number (int)
#   sample["json"]["intrinsics"]  — 3×3 camera intrinsics matrix (list of lists)

Dataset Structure

Each sample contains paired images of an indoor scene with all objects removed, leaving only the structural layout (walls, floor, ceiling):

Field	Description	Format
image	Hypersim dataset image (tonemapped renders)	1024×768 RGBA PNG
depth	Depth to the first visible structural surface	1024×768 16-bit grayscale PNG
normal	Surface normal of the first visible structural surface	1024×768 16-bit RGB PNG
layout_depth	Layout-layer depth map	1024×768 16-bit grayscale PNG
layout_normal	Layout-layer surface normal map	1024×768 16-bit RGB PNG
seen_mask	Visibility/validity mask	1024×768 8-bit grayscale PNG
json	Metadata dict with scene, camera, frame, and 3×3 intrinsics matrix	JSON