muxiddin19/d405-hand-surface-depth

D405 Hand-Surface Depth Dataset

Real-Time Multimodal Fingertip Contact Detection via Depth and Motion Fusion for Vision-Based Human-Computer Interaction

CVPR 2026

Mukhiddin Toshpulatov^1,2,4,5 · Wookey Lee² · Suan Lee³ · Geehyuk Lee¹

¹SpaceTop, SoC, KAIST    ²VoiceAI, BMSE, Inha University    ³SoCS, Semyung University    ⁴Dep. of CE, Gachon University, South Korea    ⁵Jizzakh branch of the National University of Uzbekistan

Dataset Description

A multi-user, multi-angle RGB-depth dataset captured with an Intel RealSense D405 stereo depth camera for close-range hand-surface interaction research. The dataset supports two tasks:

1. Metric depth estimation fine-tuning — close-range (7–50 cm) depth supervision for monocular models
2. Fingertip contact detection — per-fingertip contact/hover labels derived from calibrated surface plane geometry

This dataset was used to fine-tune Depth Anything V2 ViT-S, reducing depth MAE by 68%, from 12.3 mm → 3.84 mm at 25–45 cm operating range, and to train and evaluate a velocity-gated hysteresis contact detector achieving 94.2% accuracy and 94.4% F1-score.

Dataset Summary

Property	Value
Total RGB-depth pairs	53,300
Participants	15 users
Camera angles	30°, 45°, 60°, 90°
Surfaces	White desk (primary)
Resolution	640 × 480 pixels
Frame rate	30 FPS
Depth sensor	Intel RealSense D405 (active stereo)
Depth range	70–500 mm
Depth accuracy	< 0.5 mm at 350 mm
Per-frame annotations	21 MediaPipe hand landmarks + 5 fingertip metric depths
Contact labels	Per-fingertip binary contact/hover state

Splits

All frames from a given participant belong to exactly one split (participant-stratified), preventing identity leakage from hand shape, skin tone, or typing style.

Split	Frames	Participants
Train	42,640 (80%)	P01, P03, P04, P06, P07, P09–P13, P18 + supplementary
Validation	5,330 (10%)	P05, P16 + supplementary
Test	5,330 (10%)	P08, P17 + supplementary
Total	53,300	15 participants (8:1:1 ratio)

Data Structure

d405-hand-surface-depth/
├── train/
│   ├── P01/
│   │   ├── P01_angle90_white_desk_typing_074155/
│   │   │   ├── rgb/
│   │   │   │   ├── 000000.png        # BGR uint8, 640×480
│   │   │   │   ├── 000001.png
│   │   │   │   └── ...
│   │   │   ├── depth/
│   │   │   │   ├── 000000.png        # uint16, millimeters
│   │   │   │   ├── 000001.png
│   │   │   │   └── ...
│   │   │   ├── annotations/
│   │   │   │   ├── 000000.json       # Hand landmarks + fingertip depths
│   │   │   │   └── ...
│   │   │   └── metadata.json         # Session & camera info
│   │   ├── P01_angle45_white_desk_typing_094354/
│   │   │   └── ...
│   │   └── ...
│   ├── P03/
│   │   └── ...
│   └── ...
├── val/
│   ├── P13/
│   │   └── ...
│   └── P16/
│       └── ...
├── test/
│   ├── P05/
│   │   └── ...
│   └── P08/
│       └── ...
└── splits/
    ├── train.txt
    ├── val.txt
    └── test.txt

File Formats

RGB Images (rgb/*.png)

Standard BGR uint8 images (640 × 480). Compatible with OpenCV cv2.imread().

Depth Maps (depth/*.png)

16-bit unsigned integer PNGs where each pixel value represents depth in millimeters. A value of 0 indicates invalid/missing depth.

import cv2
import numpy as np

depth_mm = cv2.imread("depth/000000.png", cv2.IMREAD_UNCHANGED)  # uint16
depth_m = depth_mm.astype(np.float32) / 1000.0                    # convert to meters

Annotations (annotations/*.json)

Per-frame hand tracking results from MediaPipe:

{
  "frame_id": 17,
  "num_hands": 2,
  "hands": [
    {
      "handedness": "Right",
      "fingertip_depths_m": {
        "thumb": 0.352, "index": 0.348,
        "middle": 0.324, "ring": 0.350, "pinky": 0.357
      },
      "fingertip_pixels": {
        "thumb": [294, 326], "index": [303, 241],
        "middle": [338, 218], "ring": [368, 224], "pinky": [402, 248]
      },
      "landmarks_px": [[180, 426], [231, 418], "... (21 landmarks)"]
    }
  ]
}

Session Metadata (metadata.json)

{
  "camera": "Intel RealSense D405",
  "resolution": [640, 480],
  "fps": 30,
  "depth_scale_m": 0.0001,
  "depth_range_m": [0.07, 0.5],
  "intrinsics": {
    "fx": 434.16, "fy": 432.89,
    "cx": 323.11, "cy": 239.79,
    "width": 640, "height": 480
  },
  "session": {
    "participant": "P01",
    "angle_degrees": 90,
    "surface": "white_desk",
    "action": "typing"
  },
  "surface_calibration": {
    "plane_coefficients": [-1.24e-05, 1.63e-05, 1.0, -0.3647],
    "inlier_ratio": 0.77,
    "avg_surface_depth_m": 0.363
  }
}

Camera Setup

         [90° overhead]
              |
              |  35 cm
              |
    ──────────┼──────────  ← desk surface
   /          |          \
  / 60°     45°      30° \
 /            |            \
[cam]       [cam]        [cam]

The Intel RealSense D405 is mounted on a tripod at 25–45 cm from the desk surface. Each participant is recorded at multiple angles while typing on a printed QWERTY keyboard layout.

D405 Depth Processing Pipeline

Filter	Parameter	Value
Spatial	magnitude	2
Spatial	alpha	0.5
Spatial	delta	20
Temporal	alpha	0.4
Temporal	delta	20
Hole filling	mode	1 (farthest from around)

Contact Labels

Binary contact/hover labels are derived automatically from depth measurements using a velocity-gated hysteresis state machine:

Transition	Threshold
Hover → Contact (entry)	fingertip-to-surface ≤ 4.5 mm
Contact → Hover (exit)	fingertip-to-surface ≥ 6.0 mm
Cooldown	450 ms (~15 frames at 30 FPS)

Surface depth at each pixel is computed from RANSAC plane coefficients fitted during calibration (empty desk, 30 frames).

Intended Use

Primary Use Cases

• Monocular metric depth estimation — fine-tuning depth foundation models (e.g., Depth Anything V2, ZoeDepth, Metric3D) for close-range hand-surface scenarios
• Fingertip contact detection — training and evaluating touch/hover classifiers for vision-based virtual keyboards and touchless interfaces
• Hand-surface interaction research — studying hand pose, depth, and contact dynamics across multiple viewpoints

Out-of-Scope Use

• General-purpose depth estimation (this dataset is specialized for 7–50 cm range)
• Biometric identification from hand shape or skin tone
• Any application that requires individual participant re-identification

Results Using This Dataset

Depth Estimation (fine-tuned Depth Anything V2 ViT-S)

Metric	Pre-trained	Fine-tuned
MAE (mm)	12.3	3.84
δ1 (%)	87.2	95.96
RMSE (mm)	18.4	4.8
abs_rel	0.042	0.008

Contact Detection

Method	Accuracy (%)	F1 (%)	FPR (%)
Depth threshold only	87.3	86.1	8.7
Velocity only	89.1	88.5	6.3
Depth + velocity fusion (ours)	94.2	94.4	4.2

Ethical Considerations

• All participants provided informed consent for data collection and public release
• No personally identifiable information (face, name) is included — only hand images
• The dataset should not be used for biometric identification or surveillance purposes
• Participant IDs (P01–P18) are anonymized

Citation

@inproceedings{toshpulatov2026realtime,
  title={Real-Time Multimodal Fingertip Contact Detection via Depth and Motion
         Fusion for Vision-Based Human-Computer Interaction},
  author={Toshpulatov, Mukhiddin and Lee, Wookey and Lee, Suan and Lee, Geehyuk},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision
             and Pattern Recognition (CVPR)},
  year={2026}
}

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Acknowledgements

• Intel RealSense for the D405 depth camera SDK
• MediaPipe for hand landmark detection
• Depth Anything V2 for the base depth estimation model