README.md

---
license: cc-by-4.0
task_categories:
- depth-estimation
tags:
- robotics
- 3d
- simulation
---

# ByteCameraDepth Dataset

[Paper](https://huggingface.co/papers/2509.02530) | [Project Page](https://manipulation-as-in-simulation.github.io/) | [Code](https://github.com/ByteDance-Seed/manip-as-in-sim-suite)

ByteCameraDepth is a multi-camera depth estimation dataset containing synchronized depth, color, and auxiliary data captured from various 3D cameras. The dataset provides comprehensive depth sensing from multiple cameras in various in-door scenarios, making it ideal for developing and evaluating depth estimation algorithms.

## Dataset Overview

- **Purpose**: Multi-camera depth estimation research and benchmarking
- **Total Sessions**: 39 recording sessions
- **Unpacked Size**: ~2.7TB
- **Data Collection System**: [Multi-Camera Depth Recording System](https://github.com/Ericonaldo/depth_recording)
- **License**: CC-BY-4.0

## Quick Start

### Data Extraction

The dataset is provided as split archive files. To extract the complete dataset:

```bash
cat recorded_data.tar.part.* | tar -xvf -
```

This will create a `recorded_data` folder containing all 39 recording sessions.

### Sample Usage

To run depth inference on RGB-D camera data using the CDM (Camera Depth Models) trained with this dataset:

```bash
cd cdm
python infer.py \
    --encoder vitl \
    --model-path /path/to/model.pth \
    --rgb-image /path/to/rgb.jpg \
    --depth-image /path/to/depth.png \
    --output result.png
```

## Dataset Structure

### Archive Organization

```
recorded_data_packed/
├── recorded_data.tar.part.000
├── recorded_data.tar.part.001
├── ...
└── recorded_data.tar.part.136
```

### Extracted Data Structure

After extraction, the data is organized as follows:

```
recorded_data/
└── YYYYMMDD_HHMM/              # Timestamp-based session folder (39 sessions total)
    ├── camera_realsense_455/    # Intel RealSense D455
    │   ├── depth_000.png        # 16-bit depth images
    │   ├── color_000.png        # 8-bit color images
    │   └── ...
    ├── camera_realsense_d405/   # Intel RealSense D405
    │   ├── depth_000.png
    │   ├── color_000.png
    │   └── ...
    ├── camera_realsense_d415/   # Intel RealSense D415
    │   ├── depth_000.png
    │   ├── color_000.png
    │   └── ...
    ├── camera_realsense_d435/   # Intel RealSense D435
    │   ├── depth_000.png
    │   ├── color_000.png
    │   └── ...
    ├── camera_realsense_l515/   # Intel RealSense L515
    │   ├── depth_000.png
    │   ├── color_000.png
    │   └── ...
    ├── camera_kinect/           # Microsoft Azure Kinect
    │   ├── depth_000.png        # 16-bit depth images
    │   ├── color_000.png        # 8-bit color images
    │   ├── ir_000.png           # Infrared images
    │   └── ...
    ├── camera_zed2i_neural/     # Stereolabs ZED2i (Neural mode)
    │   ├── raw_depth_000.npy    # 32-bit float depth arrays
    │   ├── depth_000.png        # 16-bit depth images
    │   ├── color_000.png        # Color images
    │   ├── pcd_000.npy          # Point cloud data (X,Y,Z)
    │   ├── normal_000.npy       # Surface normal vectors
    │   └── ...
    ├── camera_zed2i_performance/
    ├── camera_zed2i_quality/
    ├── camera_zed2i_ultra/
    └── ...
```

## Camera Systems and Specifications

The dataset includes data collected by our [depth recording toolkit](https://github.com/Ericonaldo/depth_recording):

### Intel RealSense Cameras

- **Models**: D405, D415, D435, D455, L515
- **Output**: `depth_xxx.png` (16-bit), `color_xxx.png` (8-bit)

### Microsoft Azure Kinect

- **Depth Resolution**: Wide FOV unbinned
- **Output**: `depth_xxx.png` (16-bit), `color_xxx.png` (8-bit), `ir_xxx.png` (infrared)

### Stereolabs ZED2i

- **Depth Resolution**: 1280×720
- **Depth Modes**: 4 different modes (neural, performance, quality, ultra)
- **Output**: 
  - `raw_depth_xxx.npy` (32-bit float depth arrays)
  - `depth_xxx.png` (16-bit depth images)
  - `color_xxx.png` (8-bit color images)
  - `pcd_xxx.npy` (point cloud data)
  - `normal_xxx.npy` (surface normal vectors)

## Data Formats

### File Types and Specifications

| Data Type | Format | Bit Depth | Description |
|-----------|--------|-----------|-------------|
| Depth Images | PNG | 16-bit | Standard depth maps |
| Color Images | PNG | 8-bit RGB | Color/texture images |
| Raw Depth | NPY | 32-bit float | High-precision depth (ZED2i only) |
| Point Clouds | NPY | 32-bit float | 3D point coordinates (X,Y,Z) |
| Surface Normals | NPY | 32-bit float | Surface normal vectors |
| Infrared | PNG | 8-bit | IR images (Kinect only) |

### Depth Data
The unit of the depth data is 'mm' for most of the cameras, which means that we can obtain the 'm'-scale by dividing the raw depth by 1000.
Note that RealSense D405/L515 has different scales, which are 2500 and 10000, respectively. In other words, we should divide the raw depth by 2500 and 10000 to obtain the 'm'-scale depth.

### File Naming Convention

- Sequential numbering: `xxx` represents frame index (000, 001, 002, ...)
- Synchronized capture: Same frame numbers across cameras represent simultaneous capture
- Camera identification: Folder names clearly identify camera type and model

## 📄 Citation

If you use this dataset in your research, please cite:

```bibtex
@article{liu2025manipulation,
  title={Manipulation as in Simulation: Enabling Accurate Geometry Perception in Robots},
  author={Liu, Minghuan and Zhu, Zhengbang and Han, Xiaoshen and Hu, Peng and Lin, Haotong and 
          Li, Xinyao and Chen, Jingxiao and Xu, Jiafeng and Yang, Yichu and Lin, Yunfeng and 
          Li, Xinghang and Yu, Yong and Zhang, Weinan and Kong, Tao and Kang, Bingyi},
  journal={arXiv preprint},
  year={2025},
  url={https://huggingface.co/papers/2509.02530}
}
```

## License

This dataset is released under the CC BY 4.0 License.