muxiddin19/d405-hand-surface-depth

D405 Hand-Surface Depth Dataset

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

Paper: Real-Time Multimodal Fingertip Contact Detection via Depth and Motion Fusion for Vision-Based Human-Computer Interaction (CVPR 2026)

Authors: Mukhiddin Toshpulatov, Wookey Lee, Suan Lee, Geehyuk Lee

Key Statistics

Metric	Value
Total RGB-depth pairs	53,300
Participants	15 (anonymized P01-P18)
Camera angles	30°, 45°, 60°, 90°
Resolution	640 × 480
Frame rate	30 FPS
Depth range	70-500 mm
Depth format	uint16 PNG (millimeters)
RGB format	BGR uint8 PNG

Splits (Participant-Stratified)

Split	Frames	Participants
Train	42,954	P01, P03, P04, P06, P07, P09-P13, P18
Validation	5,040	P05, P16
Test	5,306	P08, P17

Participant-stratified splits prevent identity leakage across partitions.

Usage

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("muxiddin19/d405-hand-surface-depth")

# Access splits
train = dataset["train"]
val = dataset["validation"]
test = dataset["test"]

# Example: iterate over training samples
for sample in train:
    rgb_image = sample["rgb"]        # PIL Image (640x480 RGB)
    depth_image = sample["depth"]    # PIL Image (640x480 uint16, values in mm)
    participant = sample["participant"]
    angle = sample["angle"]
    has_contact = sample["has_contact"]
    print(f"{participant}, {angle}°, contact={has_contact}")
    break

Data Structure

Each sample contains:

• rgb: RGB image (640×480, BGR uint8 PNG)
• depth: Depth map (640×480, uint16 PNG, values in millimeters, 0 = invalid)
• participant: Anonymized participant ID (P01-P18)
• session: Recording session identifier
• frame_id: Frame number within session
• angle: Camera viewing angle in degrees (30, 45, 60, or 90)
• surface: Surface type (white_desk, wood_grain, semi_reflective)
• has_contact: Whether any fingertip is in contact state
• handedness: Detected hand (Left/Right)

Additional Annotations (in repo)

Per-frame JSON annotations in data/<participant>/<session>/annotations/ contain:

• 21 MediaPipe hand landmarks (pixel coordinates)
• 5 fingertip metric depths
• Per-fingertip binary contact/hover labels

Per-session metadata.json files contain:

• Camera intrinsics (fx, fy, cx, cy)
• Depth scale and processing filter parameters
• RANSAC surface plane coefficients

Recording Setup

• Sensor: Intel RealSense D405 (optimal range 7-50 cm, <0.5 mm accuracy at 35 cm)
• Mounting: 35 cm above desk at 45° primary angle
• Lighting: Diffuse LED, 5000K, 800 lux
• Surfaces: White desk (primary), wood grain, semi-reflective laminate
• Additional cameras: RGB cameras at 30° and 60° for multi-angle data

Benchmark Results

Fine-tuning Depth Anything V2 ViT-S on this dataset:

Metric	Pre-trained	Fine-tuned	Improvement
MAE (mm)	12.3	3.84	68% reduction
δ₁ (%)	87.2	95.96	+8.8 pp
RMSE (mm)	18.4	4.8	74% reduction

Downstream contact detection: 94.2% accuracy, 94.4% F1-score

Ethics

• All participants gave informed consent
• No PII (face/name) included — only hand images with anonymous IDs
• Not intended for biometric identification or surveillance

Links

• GitHub: Multimodal-Fingertip-Contact-Detection
• Model: vrkeyb-cvpr2026
• Project Page: Project Website

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}
}