1234shjx/FurnFloc-dataset

FurnFloc: floorplan localization data

Preview release: one HM3D-derived floor from the FloVerse-1.6K data, prepared for floorplan-based camera localization. This repository contains the original scene files and a small manifest describing their geometry. The full FurnFloc collection has 132 floors from 101 HM3D buildings; this preview contains only 00000-kfPV7w3FaU5_0 from the training split.

Download and layout

Download data/00000-kfPV7w3FaU5_0.tar.gz and extract it from the repository root:

tar -xzf data/00000-kfPV7w3FaU5_0.tar.gz

This creates:

data/00000-kfPV7w3FaU5_0/
├── floorplan.png
├── localization_frame_validity.json
├── structural_depth_metadata.json
└── traj_0/ ... traj_149/
    ├── rgb/rgb_<frame>.png
    ├── depth/depth_<frame>.png
    ├── traj_<id>.npy
    ├── traj_<id>.txt
    ├── depth.txt
    ├── localization_frame_validity.npy
    ├── structural_ray_validity.npy
    ├── structural_ray_status.npy
    └── object/object.json

The preview has 150 trajectories, 2,512 paired RGB and depth frames, and a 1600 × 1600 floorplan. RGB and depth images are 256 × 256 PNGs. Frame indices in each trajectory start at zero and align with rows in the pose array and depth.txt. scene_metadata.json records the archive SHA-256 and the world-to-floorplan transform.

Geometry and labels

• traj_<id>.npy and its text counterpart contain rows [x, y, look_x, look_y] in world metres. Heading is atan2(look_y-y, look_x-x).
• The world-to-map transform is in scene_metadata.json: pixel_x = x * scale_px_per_m + offset_x_px, pixel_y = y * scale_px_per_m + offset_y_px. Pixel origin is at the upper left, with both axes increasing right/down.
• depth/depth_<frame>.png is an 8-bit Euclidean ray-range image. Decode metres as uint8 / 25.5. Zero denotes invalid depth; 255 is clipped or censored at 10 m.
• depth.txt contains 40 projective camera-z wall distances per frame, one row per RGB frame. The 40 horizontal sample columns and angle convention are in structural_depth_metadata.json. It describes floorplan structural depth, which differs from the visible depth PNG when furniture occludes a wall.
• localization_frame_validity.npy marks frames eligible for map localization. structural_ray_validity.npy and structural_ray_status.npy describe ray-level map support. The accompanying JSON files define policy and status counts. Keep the original frame indices even when filtering invalid frames.
• object/object.json contains trajectory-level object annotations inherited from the source data.

Example:

import numpy as np
from PIL import Image

base = "data/00000-kfPV7w3FaU5_0/traj_0"
rgb = np.asarray(Image.open(f"{base}/rgb/rgb_0.png"))
depth_range_m = np.asarray(Image.open(f"{base}/depth/depth_0.png"), dtype=np.float32) / 25.5
poses = np.load(f"{base}/traj_0.npy", allow_pickle=False)
eligible = np.load(f"{base}/localization_frame_validity.npy", allow_pickle=False)

Provenance and use

FurnFloc is a curated localization derivative of the HM3D portion of FloVerse-1.6K, which in turn uses the Habitat-Matterport 3D Research Dataset. Please cite FloVerse and HM3D when using these data. The underlying Matterport data and information derived from it are subject to the Matterport academic-use agreement, including its use and downstream distribution conditions. The source data's terms apply; this card does not grant a new license to the RGB-D images, floorplan, or derived labels.

This preview is for research on floorplan localization, depth estimation, and robustness to furniture occlusion. It contains rendered views of real captured interiors via HM3D. Do not use the scene ID or visual content to identify a location or owner.

Citation

@inproceedings{floverse2026,
  title={FloVerse: Floor Plan-Guided Multi-Modal Navigation},
  author={Huang, Weiqi and Dong, Shuangyi and Li, Jiaxin and Guo, Yifei and Wang, Zan and Liang, Wei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}

@inproceedings{ramakrishnan2021hm3d,
  title={Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI},
  author={Ramakrishnan, Santhosh Kumar and others},
  booktitle={NeurIPS Datasets and Benchmarks Track},
  year={2021}
}

A FurnFloc-specific citation can be added after its associated paper is public.