Datasets:

jagennath-hari
/

nyuv2

language:
  - en
pretty_name: NYUv2
tags:
  - robotics
license: mit
task_categories:
  - depth-estimation
  - image-segmentation
  - image-feature-extraction
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: rgb
      dtype: image
    - name: depth
      dtype: image
    - name: semantic
      dtype: image
    - name: instance
      dtype: image
  splits:
    - name: train
      num_bytes: 2237596040.926
      num_examples: 1014
    - name: val
      num_bytes: 639427936
      num_examples: 290
    - name: test
      num_bytes: 320361261
      num_examples: 145
  download_size: 1284641786
  dataset_size: 3197385237.926

NYUv2

This is an unofficial and preprocessed version of NYU Depth Dataset V2 made available for easier integration with modern ML workflows. The dataset was converted from the original .mat format into a split structure with embedded RGB images, depth maps, semantic masks, and instance masks in Hugging Face-compatible format.

📸 Sample Visualization

RGB

Depth (Jet colormap)

Semantic Mask

Dataset Description

Homepage: NYU Depth Dataset V2
Paper: Indoor Segmentation and Support Inference from RGBD Images

NYUv2 is a benchmark RGB-D dataset widely used for scene understanding tasks such as:

Indoor semantic segmentation
Depth estimation
Instance segmentation

This version has been preprocessed to include aligned:

Undistorted RGB images (.png)
Depth maps in millimeters (.tiff, uint16)
Semantic masks (.tiff, scaled uint16)
Instance masks (.tiff, scaled uint16)

Each sample is annotated with a consistent id and split across train/val/test.

🧾 Dataset Metadata

Additional files included:

camera_params.json — camera intrinsics and distortion
class_names.json — mapping from class IDs to human-readable names
scaling_factors.json — used for metric depth and label/mask de-scaling during training

🚀 How to Use

You can load the dataset using the datasets library:

from datasets import load_dataset

dataset = load_dataset("jagennath-hari/nyuv2", split="train")
sample = dataset[0]

# Access fields
rgb = sample["rgb"]
depth = sample["depth"]
semantic = sample["semantic"]
instance = sample["instance"]

🔄 Recover Original Values from TIFF Images

The dataset uses .tiff format for all dense outputs to preserve precision and visual compatibility. Here’s how to revert them back to their original values:

from datasets import load_dataset
from huggingface_hub import snapshot_download
from PIL import Image
import numpy as np
import json
import os

# Load sample
dataset = load_dataset("jagennath-hari/nyuv2", split="train")
sample = dataset[0]

# Download and load scaling metadata
local_dir = snapshot_download(
    repo_id="jagennath-hari/nyuv2",
    repo_type="dataset",
    allow_patterns="scaling_factors.json"
)
with open(os.path.join(local_dir, "scaling_factors.json")) as f:
    scale = json.load(f)

depth_scale = scale["depth_scale"]
label_max = scale["label_max_value"]
instance_max = scale["instance_max_value"]

# === Unscale depth (mm → m)
depth_img = np.array(sample["depth"])
depth_m = depth_img.astype(np.float32) / depth_scale

# === Unscale semantic mask
sem_scaled = np.array(sample["semantic"])
semantic_labels = np.round(
    sem_scaled.astype(np.float32) * (label_max / 65535.0)
).astype(np.uint16)

# === Unscale instance mask
inst_scaled = np.array(sample["instance"])
instance_ids = np.round(
    inst_scaled.astype(np.float32) * (instance_max / 65535.0)
).astype(np.uint16)

📐 Scaling Factors Summary

Field	Stored As	Original Format	Scaling Method	Undo Formula
`depth`	`uint16`, mm	`float32`, meters	multiplied by `depth_scale`	`depth / depth_scale`
`semantic`	`uint16`, scaled	`uint16` class IDs	scaled by `65535 / label_max`	`round(mask * (label_max / 65535.0))`
`instance`	`uint16`, scaled	`uint16` instance IDs	scaled by `65535 / instance_max`	`round(mask * (instance_max / 65535.0))`

📄 Citation

If you use this dataset, please cite the original authors:

@inproceedings{Silberman:ECCV12,
  author    = {Nathan Silberman, Derek Hoiem, Pushmeet Kohli and Rob Fergus},
  title     = {Indoor Segmentation and Support Inference from RGBD Images},
  booktitle = {ECCV},
  year      = {2012}
}