📄 Technical Report | 📄 arXiv | 🌐 Project Page | 💻 Code | 🤗 Hugging Face | 🤖 ModelScope ｜ 🤖 Video

LingBot-Depth transforms incomplete and noisy depth sensor data into high-quality, metric-accurate 3D measurements. By jointly aligning RGB appearance and depth geometry in a unified latent space, our model serves as a powerful spatial perception foundation for robot learning and 3D vision applications.

Our approach refines raw sensor depth into clean, complete measurements, enabling:

• Depth Completion & Refinement: Fills missing regions with metric accuracy and improved quality
• Scene Reconstruction: High-fidelity indoor mapping with a strong depth prior
• 4D Point Tracking: Accurate dynamic tracking in metric space for robot learning
• Dexterous Manipulation: Robust grasping with precise geometric understanding

We provide pretrained models for different scenarios:

• The curated 3M RGB-D dataset will be released upon completion of the necessary licensing and approval procedures.
• Expected release: mid-March 2026.

• Python ≥ 3.9 • PyTorch ≥ 2.0.0 • CUDA-capable GPU (recommended)

git clone https://github.com/robbyant/lingbot-depth
cd lingbot-depth

# Install the package (use 'python -m pip' to ensure correct environment)
conda create -n lingbot-depth python=3.9
conda activate lingbot-depth
python -m pip install -e .

Inference:

import torch
import cv2
import numpy as np
from mdm.model.v2 import MDMModel

# Load model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = MDMModel.from_pretrained('robbyant/lingbot-depth-pretrain-vitl-14').to(device)

# Load and prepare inputs
image = cv2.cvtColor(cv2.imread('examples/0/rgb.png'), cv2.COLOR_BGR2RGB)
h, w = image.shape[:2]
image = torch.tensor(image / 255, dtype=torch.float32, device=device).permute(2, 0, 1)[None]

depth = cv2.imread('examples/0/raw_depth.png', cv2.IMREAD_UNCHANGED).astype(np.float32) / 1000.0
depth = torch.tensor(depth, dtype=torch.float32, device=device)[None]

intrinsics = np.loadtxt('examples/0/intrinsics.txt')
intrinsics[0] /= w  # Normalize fx and cx by width
intrinsics[1] /= h  # Normalize fy and cy by height
intrinsics = torch.tensor(intrinsics, dtype=torch.float32, device=device)[None]

# Run inference
output = model.infer(
    image,
    depth_in=depth,
    intrinsics=intrinsics)

depth_pred = output['depth']  # Refined depth map
points = output['points']      # 3D point cloud

Run example:

The model is automatically downloaded from Hugging Face on first run (no manual download needed):

# Basic usage - processes example 0
python example.py

# Use a different example (0-7 available)
python example.py --example 1

# Use depth completion optimized model
python example.py --model robbyant/lingbot-depth-postrain-dc-vitl14

# Custom output directory
python example.py --output my_results

# See all options
python example.py --help

This processes the example data and saves results to result/ (or custom directory):

result/
├── rgb.png                 # Input RGB image
├── depth_input.npy        # Input depth (float32, meters)
├── depth_refined.npy      # Refined depth (float32, meters)
├── depth_input.png        # Input depth visualization
├── depth_refined.png      # Refined depth visualization
├── depth_comparison.png   # Side-by-side comparison
└── point_cloud.ply       # 3D point cloud

Available examples: 8 example scenes (0-7) included in examples/ directory.

We introduce a masked depth modeling approach that learns robust RGB-D representations through self-supervised learning. The model employs a Vision Transformer encoder with specialized depth-aware attention mechanisms to jointly process RGB and depth inputs.

Depth-aware attention visualization. Visualizing attention from depth queries (Q1–Q3, marked with ⋆) to RGB tokens in two scenes: (a) aquarium and (b) indoor shelf. Each row shows masked input depth, attention weights on RGB, and refined output. Different queries attend to spatially corresponding regions, demonstrating cross-modal alignment.

Key Innovations:

• Masked Depth Modeling: Self-supervised pre-training via depth reconstruction
• Cross-Modal Attention: Joint RGB-Depth alignment in unified latent space
• Metric-Scale Preservation: Maintains real-world measurements for downstream tasks

Our model is trained on a large-scale diverse dataset combining real-world and simulated RGB-D captures:

Training dataset. 2M real-world and 1M simulated samples spanning diverse indoor environments (top). Representative RGB-D inputs with ground truth depth (bottom).

Dataset Composition:

• Real Captures: 2M samples from residential, office, and commercial environments
• Simulated Data: 1M photo-realistic renders with perfect ground truth
• Modalities: RGB images, raw depth, refined ground truth depth
• Diversity: Multiple sensor types, lighting conditions, and scene complexities

LingBot-Depth provides metric-accurate 3D geometry essential for tracking dynamic targets:

4D point tracking. Robust tracking in gym environments with dynamic human motion. Top: query point selection. Middle: 3D tracking on deforming geometry. Bottom: refined depth maps. Demonstrated on scooter, rowing machine, gym equipment, and pull-up bar.

High-quality geometric understanding enables reliable robotic grasping across diverse objects and materials:

Dexterous grasping. Robust manipulation enabled by refined depth. Top: point cloud reconstruction. Bottom: successful grasps on steel cup, glass cup, storage box, and toy car.

We developed a scalable RGB-D capture system for large-scale data collection:

RGB-D capture system. Multi-sensor setup with Intel RealSense, Orbbec Gemini, and Azure Kinect for scalable real-world data collection.

• Encoder: Vision Transformer (Large) with RGB-D fusion
• Decoder: Multi-scale feature pyramid with specialized heads
• Heads: Depth regression
• Training: Masked depth modeling with reconstruction objective

RGB Image:

• Shape: [B, 3, H, W] normalized to [0, 1]
• Format: PyTorch tensor, float32

Depth Map:

• Shape: [B, H, W]
• Unit: Meters (configurable via scale parameter)
• Invalid regions: 0 or NaN

Camera Intrinsics:

• Shape: [B, 3, 3]
• Normalized format: fx'=fx/W, fy'=fy/H, cx'=cx/W, cy'=cy/H
• Example:

  [[fx/W,   0,   cx/W],
   [  0,  fy/H,  cy/H],
   [  0,    0,    1  ]]
  

The model returns a dictionary:

{
    'depth': torch.Tensor,   # Refined depth [B, H, W]
    'points': torch.Tensor,  # Point cloud [B, H, W, 3] in camera space
}

model.infer(
    image,                                   # RGB tensor [B, 3, H, W]
    depth_in=None,                           # Input depth [B, H, W]
    use_fp16=True,                           # Mixed precision inference
    intrinsics=None,                         # Camera intrinsics [B, 3, 3]
)

If you find this work useful for your research, please cite:

@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}
}

Please also consider citing DINOv2, which serves as our backbone:

@article{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and others},
  journal={Transactions on Machine Learning Research},
  year={2024}
}

This project is released under the Apache License 2.0. See LICENSE file for details.

This work builds upon several excellent open-source projects:

• DINOv2 - Self-supervised vision transformer backbone
• Masked Autoencoders - Self-supervised learning framework
• The broader open-source computer vision and robotics communities

For questions, discussions, or collaborations:

• Issues: Open an issue on GitHub
• Email: Contact Dr. Bin Tan ([email protected]) or Dr. Nan Xue ([email protected])