Datasets:

harpreetsahota
/

lingbot-depth-subset

annotations_creators: []
language: en
license: apache-2.0
size_categories:
  - 10K<n<100K
task_categories:
  - depth-estimation
task_ids: []
pretty_name: lingbot-depth-subset
tags:
  - 3d
  - depth-completion
  - depth-estimation
  - rgb-d
  - robotics
  - fiftyone
  - group
description: >-
  A representative, recon-sampled subset of robbyant/mdm_depth (the
  LingBot-Depth training corpus), reformatted into FiftyOne's grouped-dataset
  structure with gt_depth/raw_depth as fo.Heatmap fields.
dataset_summary: >

  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 13,149
  samples (10,207 groups) spanning 3 sub-collections (RobbyReal, RobbyVla,
  RobbySim).


  ## Installation


  If you haven't already, install FiftyOne:


  ```bash

  pip install -U fiftyone

  ```


  ## Usage


  ```python

  import fiftyone as fo

  from fiftyone.utils.huggingface import load_from_hub


  # Load the dataset

  # Note: other available arguments include 'max_samples', etc

  dataset = load_from_hub("harpreetsahota/lingbot-depth-subset")


  # Launch the App

  session = fo.launch_app(dataset)

  ```

Dataset Card for lingbot-depth-subset

This is a FiftyOne dataset with 13,149 samples (10,207 groups) spanning 3 sub-collections (RobbyReal, RobbyVla, RobbySim).

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/lingbot-depth-subset")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

This is a representative, recon-sampled subset of robbyant/mdm_depth (the LingBot-Depth RGB-D training corpus), reformatted into FiftyOne's native grouped-dataset structure. The source dataset is a self-curated 3.02M-sample RGB-D corpus (2.71–3.03 TB) used to pretrain LingBot-Depth, a Masked Depth Modeling (MDM) Vision Transformer that treats missing/invalid depth-sensor pixels as "natural masks" and learns to reconstruct dense, metric depth from RGB context plus the remaining valid depth readings.

Because the source repo ships as ~20 multi-hundred-GB .tar.zst archives, this subset was produced by streaming each archive (curl | zstd -d | tar -x) and interrupting the stream early, capturing whichever scenes/sequences appear first in each archive — not a uniform random sample across the full 3M-sample corpus. It exists to validate FiftyOne import tooling and enable exploratory analysis of the dataset's structure, formats, and known quirks (see "Parsing decisions" below) without downloading the full multi-terabyte dataset.

Curated by: Original data: the Robbyant team at Ant Group (paper authors: Bin Tan, Changjiang Sun, Xiage Qin, Hanat Adai, Zelin Fu, Tianxiang Zhou, Han Zhang, Yinghao Xu, Xing Zhu, Yujun Shen, Nan Xue). This FiftyOne-formatted subset: harpreetsahota.
Funded by: [More Information Needed]
Shared by: harpreetsahota (this FiftyOne conversion); original data shared by robbyant (Ant Group)
Language(s): N/A — image and depth-sensor data, no natural language content
License: apache-2.0. Caveat: the upstream robbyant/mdm_depth dataset card itself is inconsistent — its YAML frontmatter states apache-2.0 but its rendered card body explicitly states "License: CC BY-NC-SA 4.0" (non-commercial, share-alike). This card follows the frontmatter-stated license, but you should verify licensing terms directly with the original authors before any commercial use.

Dataset Sources

Repository: https://huggingface.co/datasets/harpreetsahota/lingbot-depth-subset (this dataset); https://huggingface.co/datasets/robbyant/mdm_depth (source data); https://github.com/robbyant/lingbot-depth (code)
Paper: Masked Depth Modeling for Spatial Perception (arXiv:2601.17895)
Demo: [More Information Needed]

Uses

Direct Use

Prototyping and QA of FiftyOne import pipelines for large, multi-format RGB-D datasets packaged as compressed archives.
Exploratory analysis of RGB-D depth-completion data: comparing raw sensor depth against pseudo-ground-truth depth, visualizing missing-depth patterns on reflective/transparent surfaces, inspecting per-sub-dataset conventions (file naming, invalid-pixel sentinels, camera intrinsics).
Small-scale experimentation with masked depth modeling / depth completion methods where a full multi-terabyte download is impractical.
Studying multi-view capture structure (e.g. the RobbySim object_view 10-camera rig) via FiftyOne's grouped-dataset slices.

Out-of-Scope Use

Not a substitute for the full robbyant/mdm_depth corpus (3,019,200 samples) or the paper's full 10M-sample MDM training corpus (which additionally folds in 7 external open datasets: ClearGrasp, Hypersim, ADT, ArkitScenes, TartanAir, ScanNet++, Taskonomy). Do not use this subset to reproduce the paper's reported training results.
Not a uniformly random sample — coverage of scenes, sensors, and robots (e.g. only the franka robot, not ur7e) is incomplete and biased toward whatever appeared first in each source archive.
Not vetted for commercial use given the license ambiguity noted above.

Dataset Structure

This is a grouped FiftyOne dataset (dataset.media_type == "group"): each native FiftyOne "group" corresponds to one physical capture instant (a scene/sequence + frame index), and each populated "slice" within a group corresponds to one sensor/camera/side that captured that instant. len(dataset) reports only the default group slice's count (right_realsense405, 4,636 samples) — use len(dataset.select_group_slices()) for the true total of 13,149 samples in 10,207 groups.

Sub-collections (accessible as saved views)

Saved view	Sub-dataset	Samples	Groups	Group-size distribution
`RobbyReal`	Real-world indoor RGB-D captures	1,443	1,443	all size 1 (no multi-sensor overlap observed in this slice)
`RobbyVla`	Real-world VLA robot manipulation (Franka arm)	9,143	8,264	7,385 size-1, 879 groups with genuine `left`+`right` stereo matches
`RobbySim`	Simulated renders (`object_view`, `rrt_view`, `val_view`)	2,563	500	ranges up to size 10 (the `object_view` 10-camera rig groups correctly)

RobbySim further splits by the split field (train for object_view/rrt_view, val for val_view, matching the source RobbySimVal split) and the view field (object_view, rrt_view, val_view).

Group slices (sensor/camera/side identifiers)

Slice names vary by sub-dataset and are not synchronized across all of them — see "Parsing decisions" below:

RobbyReal: sensor identifiers, e.g. orbbec_335_CP026530001N (only one sensor type appears in this particular slice; the source supports 5: orbbec_335, orbbec_335L, realsense_D415, realsense_D435, realsense_D455)
RobbyVla: left_realsense405, right_realsense405
RobbySim (object_view/rrt_view/val_view): camera IDs 00–19 (which IDs are populated depends on the view/scene)

Additional saved dynamic-group "video" views — orbbec_335_CP026530001N_as_video, left_realsense405_as_video, right_realsense405_as_video — group each continuous sensor stream by scene/sequence and order by frame_index, so the FiftyOne App can render them with video-scrubber controls in the sample modal (enable "Render frames as video" in display options). These are only provided for RobbyReal/RobbyVla, since RobbySim's frame indices are not temporally meaningful (see below).

Fields

Field	FiftyOne type	Description
`filepath`	`StringField`	Path to the RGB image (verbatim from source `color`/`_rgb`/`_left` files)
`group_field`	`Group`	Native FiftyOne group field; links samples captured at the same scene/frame instant across slices
`sub_dataset`	`StringField`	One of `RobbyReal`, `RobbyVla`, `RobbySim`
`split`	`StringField`	`train` or `val` (derived: `val` only for `RobbySim`'s `val_view` samples, matching source `RobbySimVal`)
`view`	`StringField`	For `RobbySim` only: `object_view`, `rrt_view`, or `val_view` (verbatim from source directory structure)
`scene_id`	`StringField`	Scene/scan identifier (`RobbyReal`/`RobbySim`) or sequence identifier (`RobbyVla`), verbatim from source directory names
`sensor_id`	`StringField`	`RobbyReal` only: physical sensor identifier (verbatim)
`robot`	`StringField`	`RobbyVla` only: robot platform (`franka`; source also defines `ur7e`, not present in this slice)
`stereo_side`	`StringField`	`RobbyVla` only: `left_realsense405` or `right_realsense405`
`cam_id`	`StringField`	Convenience copy of the slice name (sensor/cam/side), populated for every sample regardless of sub-dataset
`frame_index`	`IntField`	Frame number parsed from the source filename. Temporally meaningful (frame N follows frame N-1) for `RobbyReal` and `RobbyVla`; for `RobbySim` it is closer to a per-camera enumeration/render-instance ID than a time axis (see "Parsing decisions")
`fx`, `fy`, `cx`, `cy`	`FloatField`	Pinhole camera intrinsics, parsed from the source's `intrinsic(s).txt` (per-sensor for `RobbyReal`/`RobbyVla`/`RobbySimVal`, one shared file for `RobbySim` object_view/rrt_view)
`gt_depth`	`Heatmap`	Pseudo-ground-truth depth (stereo-derived for real captures, perfect render for synthetic), `map_path` pointing directly at the original 16-bit PNG (mm), with a per-sample `range` — see below
`raw_depth`	`Heatmap`	Raw/uncorrected sensor (or simulated-sensor) depth before post-processing, same encoding as `gt_depth`

dataset.info is currently empty — no dataset-level provenance/calibration metadata is stored beyond the per-sample intrinsics fields above.

Parsing decisions

Archives, not loose files. The source repo ships as large .tar.zst archives (some split into .partNN chunks) rather than individual files; this subset was produced by streaming each archive and interrupting extraction early (see "Dataset Description" above), not by downloading complete archives.
Three incompatible filename conventions. RobbySim's object_view/rrt_view use NNNN_camK_left.jpg / _depth.png / _rmd2c.png; val_view instead uses NNNN_camK_rgb.left.jpg / _depth_left.png / _rawdepth.left.png; RobbyReal/RobbyVla use plain numeric filenames inside separate color//gtdepth//rawdepth/ folders. Each is parsed with its own regex/glob in the ingest script.
Invalid-pixel sentinel differs by split. 0 is the missing/invalid-depth marker everywhere except RobbySim's val_view gt_depth, which additionally uses 65535 (uint16 max) as an invalid marker (~2% of pixels in samples checked). The ingest script excludes both 0 and, where applicable, 65535 when computing valid-pixel statistics.
Heatmap range is percentile-based, not raw min/max. Raw min/max (as used in FiftyOne's own NYU Depth V2 tutorial) is not robust here: raw sensor depth on reflective/transparent surfaces produces rare far-outlier pixels (observed: one frame's valid depth spanned 58–3453mm, but the 99th percentile was only 292mm) that would otherwise crush 99% of legitimate values into a sliver of the color scale. Each Heatmap's range is instead set to the [1st, 99th] percentile of that sample's valid (non-sentinel) pixels.
RobbySim's object_view genuinely is a synchronized multi-view rig — pairwise frame_index overlap (Jaccard) across its 10 fixed cameras in one scene ranged 0.52–0.93 (mean ≈0.73), confirming frame_index is largely a shared render-instance ID across cameras, with some per-camera dropout. RobbyVla's left/right RealSense405 streams, despite being physically co-mounted, are not frame-synchronized (mean Jaccard ≈0.14 across 30 sequences, many with zero overlap) — most RobbyVla groups therefore end up single-slice, with 879 groups having a genuine match.
frame_index is not a time axis for RobbySim. RobbyReal and RobbyVla frame indices come from genuinely continuous capture (verified by inspecting consecutive frames for smooth camera/gripper motion). RobbySim's object_view frames at a fixed camera pose instead vary in scene texture/material between indices (independent renders reusing a camera pose), and rrt_view/val_view camera-id folders show negligible inter-frame motion — consistent with "rrt" (Rapidly-exploring Random Tree) denoting randomized viewpoint sampling rather than a trajectory over time.
5 known-truncated files were excluded. Because source directories were populated via interrupted streaming downloads, 5 files (out of 40,364 checked) were confirmed truncated/corrupt; their entire frame-triplet (RGB + gt_depth + raw_depth) was dropped from this subset.
A more detailed recon writeup (media/label inventory, per-sub-dataset statistics, discrepancies found vs. the source card) is available in this dataset's repository as DATASET_INSPECTION.md.

Dataset Creation

Curation Rationale

This subset was curated to (a) validate a FiftyOne import pipeline against a large, heterogeneously-formatted, multi-terabyte RGB-D dataset without downloading it in full, and (b) provide a representative, browsable sample of all of the source dataset's sub-collections, camera configurations, and known data-quality quirks for exploratory analysis.

Source Data

Data Collection and Processing

Per the source paper, the original 3.02M-sample corpus was collected via two parallel pipelines:

Synthetic pipeline: RGB, perfect rendered depth, and stereo image pairs (with speckle patterns) rendered in Blender from self-hosted 3D assets, with the stereo pairs processed by a semi-global matching (SGM) algorithm to synthesize realistic sensor-like depth artifacts (RobbySim).
Real-world pipeline: a custom, modular, 3D-printed multi-camera capture rig compatible with several commercial RGB-D cameras (Orbbec Gemini 335/335L, Intel RealSense D415/D435/D455/405, ZED-mini), deployed across residential, office, commercial, public, and outdoor scenes (RobbyReal), and separately during Franka-arm/UR7e-arm robot manipulation episodes (RobbyVla). Pseudo-ground-truth depth for real captures was computed from left-right IR stereo pairs with a left-right consistency check to filter unreliable pixels.

For this subset specifically, media was obtained by streaming the source .tar.zst archives directly from the Hugging Face Hub and interrupting extraction after a few hundred MB to a few GB per archive, then validating every extracted image for corruption (see "Parsing decisions").

Who are the source data producers?

The Robbyant team (Ant Group). Real-world data was captured with a custom capture rig operated by the data collection team across a range of environments (see the source paper's Table 1: residential, office/study, commercial/service, public, and outdoor spaces); robot manipulation data was collected using Franka and UR7e robot arms.

Annotations

Annotation process

There is no manual annotation. The "labels" in this dataset are the paired depth maps themselves: gt_depth is derived algorithmically (stereo matching + left-right consistency filtering for real captures; direct rendering for synthetic scenes), and raw_depth is the depth sensor's (or simulated sensor's) uncorrected output.

Who are the annotators?

N/A — no human annotators were involved in producing the depth labels.

Personal and Sensitive Information

The RobbyReal real-world captures span a range of indoor (and some outdoor) environments, including public and semi-public spaces (e.g. hospitals/clinics, retail stores, restaurants, per the source paper). No explicit statement is made by the original authors regarding faces or other personally identifiable information appearing in these scenes. [More Information Needed]

Citation

BibTeX:

@article{lingbot-depth2026,
  title={Masked Depth Modeling for Spatial Perception},
  author={Tan, Bin and Sun, Changjiang and Qin, Xiage and Adai, Hanat and Fu, Zelin and Zhou, Tianxiang and Zhang, Han and Xu, Yinghao and Zhu, Xing and Shen, Yujun and Xue, Nan},
  journal={arXiv preprint arXiv:2601.17895},
  year={2026}
}

APA:

Tan, B., Sun, C., Qin, X., Adai, H., Fu, Z., Zhou, T., Zhang, H., Xu, Y., Zhu, X., Shen, Y., & Xue, N. (2026). Masked Depth Modeling for Spatial Perception. arXiv preprint arXiv:2601.17895.

More Information

This is a partial, non-randomly-sampled subset (13,149 of the source's 3,019,200 samples) intended for tooling validation and exploratory analysis, not for reproducing the source paper's training results. See DATASET_INSPECTION.md in this repository for the full dataset recon report, including per-sub-dataset media/label inventories, quantified frame-synchronization statistics, and discrepancies found against the source dataset card.

Dataset Card Authors

harpreetsahota (FiftyOne conversion and this card); original dataset by the Robbyant team (Ant Group).

Dataset Card Contact

[More Information Needed]