DeepStream Pipeline
Hardware-accelerated processing via GStreamer/DeepStream. Runs TensorRT YOLO engines directly on GPU, passing frames through appsink for low-copy Python processing.
v3.0.1 Stable Release
We Give Robots
Vision
The open-source vision framework for edge devices. Runs DeepStream pipelines, YOLO detection, and world models at 60 FPS on NVIDIA Jetson, Intel NPU, and Hailo.
openeyes-engine - bash
$
[INFO] Initializing OpenEyes Vision Engine v2.6.0
[SYSTEM] Hardware detected: Jetson Orin Nano (8GB)
[SYSTEM] CUDA Available: True | TensorRT: True
[INFO] Loading YOLO TensorRT engine... DONE
[INFO] Initializing MediaPipe FaceMesh... DONE
[INFO] Initializing MediaPipe Hands... DONE
[INFO] Starting DeepStream pipeline via appsink...
FPS: 60 | Obj: 3 | Face: 1 | Hand: thumbs_up | Pose: 1
FPS: 60 | Obj: 4 | Face: 1 | Hand: open_palm | Pose: 1
FPS: 59 | Obj: 4 | Face: 1 | Hand: open_palm | Pose: 1
Core Architecture
Multi-Modal Inference
Simultaneous FaceMesh, hand gesture recognition, and body pose estimation running alongside primary object detection.
ROS2 Native
Publishes telemetry across dedicated topics such as detections, depth, and pose with a multithreaded executor strategy.
World Models
Predictive planning with LeWM and V-JEPA style world models for temporal awareness, safety checks, and forward simulation.
Performance Benchmarks
Tested on NVIDIA Jetson Orin Nano (8GB) in MAXN mode.
| Configuration | Models Active | Frame Rate | Latency |
|---|---|---|---|
| Detection Only (INT8) | YOLO TensorRT | 60 FPS | 16ms |
| Minimal Pipeline | Detection + Depth + Tracking | 35-40 FPS | 28ms |
| Full Pipeline | Detection + Face + Gesture + Pose | 25-30 FPS | 38ms |
| World Model Planning | LeWM (Inference) | 200 Hz | 5ms |