Skip to main content

Getting Started with sDepth

Overview

sDepth leverages cutting-edge neural networks to generate high-fidelity depth maps from stereo images, consistently delivering reliable depth estimation even in demanding conditions. Compared to traditional methods, sDepth achieves superior accuracy in texture-poor regions and low-light environments. These advantages make it ideal for autonomous robotics, Augmented Reality (AR), and large-scale 3D reconstruction applications where robust depth perception is critical.

sDepth SDK provides a suite of AI-driven depth modes, enabling you to fine-tune depth sensing configurations to meet your specific application requirements. Each mode offers a unique trade-off between spatial fidelity, sensing range, and computational efficiency, allowing you to prioritize accuracy, extend range, improve processing speed, or achieve a tailored balance based on your operational needs.

Key Features

  • AI-Powered Depth Estimation: Neural network-based depth estimation superior to traditional methods
  • Robust Performance: Reliable accuracy in texture-poor regions and low-light environments
  • Flexible Modes: Multiple depth modes with configurable trade-offs
  • Wide Applications: Autonomous robots, AR, 3D reconstruction

Supported Platforms

  • NVIDIA Jetson AGX Orin
  • NVIDIA Jetson AGX Thor

Real-time Inference Performance

The following performance data was measured under indoor conditions using randomly collected scenes with real-time inference exceeding 3 minutes. Values represent average resource utilization and frame rates.

note

On the Jetson AGX Thor (128GB, Jetpack 7.0), camera output is limited to 30FPS, which caps the model inference speed at 30FPS.

Performance on Embedded Devices

CameraEdge DeviceInterface BoardFPSCPU Avg (%)GPU Avg (%)
S36Jetson AGX Orin (32GB, Jetpack 6.2)10-ch GMSL2438.4593.43
S36Jetson AGX Orin (32GB, Jetpack 6.2)10-ch GMSL1015.4246.73
S36Jetson AGX Thor (128GB, Jetpack 7.0)8-ch GMSL3018.3536.67
S36Jetson AGX Thor (128GB, Jetpack 7.0)8-ch GMSL107.1715.24

Vertical Ranging Accuracy

Testing was conducted in an indoor environment using a laser rangefinder. Each depth distance was measured 10 times and compared against actual values.

note

All tests were performed on Jetson AGX Orin (32GB, Jetpack 6.2) with S36 camera at 24FPS in indoor conditions.


Display Settings

After running the SDK, you can switch between display modes using keyboard keys. Mode switching does not affect accuracy, only visual display effects.

Visualization Page Switching

KeyFunction
q / ESCExit depth mode
d / DSwitch between depth map / disparity map

Color Mapping

KeyModeDescription
1Linear Normalization + TURBOSuitable for short range
2Power Function Normalization + TURBOEnhanced color levels for long range

Parameter Adjustment (Mode 2 Only)

KeyFunction
+ / =Increase gamma value (reduces distant layer sensation)
-Decrease gamma value (enhances distant depth perception)

sDepth SDK

The sDepth SDK provides core capabilities including calibration parameter loading and depth map generation, supporting both offline testing and real-time camera acquisition scenarios. It significantly reduces the barrier to developing and integrating binocular depth algorithms.

SDK Download

ResourceLink
sDepth SDKThe SDK download link is currently under maintenance and will be updated soon
caution

The SDK download link is currently under maintenance and will be updated soon.

File Structure

sDepth/
├── CMakeLists.txt
├── README.md
├── config/
│ ├── stereo_calibration.json
│ └── sDepth.enc
├── include/
│ ├── opencv4/
│ └── SG_StereoDepth_api.h
├── lib/
│ ├── opencv/
│ └── libsDepth.so
└── samples/
└── main.cpp

Quick Start

1. Install Dependencies

sudo apt install -y cmake build-essential v4l-utils nvidia-jetpack

2. Build

cd sDepth
mkdir build
cd build
cmake ..
make -j$(nproc)

3. Run

The executable sDepth_Demo is generated in the sDepth/bin directory.


Parameter Reference

ParameterDescriptionRequired
-modeSDK mode: 0=offline (stereo images to depth), 1=online YUV preview, 2=online RAW previewYes
-leftOffline: left image file path; Online: left camera device numberYes
-rightOffline: right image file path; Online: right camera device numberYes
-widthInput image/stream width in pixels (must match input resolution)Yes
-heightInput image/stream height in pixels (must match input resolution)Yes
-keyModel license fileYes
-encDepth model file (default: sDepth/config/sDepth.enc)No
-jsonCamera calibration file (default: sDepth/config/stereo_calibration.json)No

Usage Examples

Offline Mode

Input stereo images and output depth map:

./sDepth_Demo -mode 0 -left left.jpg -right right.jpg -width 1920 -height 1080 -key <keyfile_name>

Online Mode

Real-time depth video stream preview:

./sDepth_Demo -mode 1 -left /dev/video0 -right /dev/video1 -width 1920 -height 1080 -key <keyfile_name>