A full NVIDIA Jetson GPU stack: IMX708 CSI capture, CUDA, cuDNN, TensorRT, GPU PyTorch, source built torchvision, YOLO pose state and a Three.js spatial renderer.
NVIDIA Jetson Orin Nano Super · TensorRT · Spatial computing
WISP Lab
An NVIDIA Jetson vision system that turns physical surfaces into responsive interfaces.
- ROLE
- Solo builder
- STATE
- Active Jetson lab prototype
CSI INPUTIMX708 CAMERA
NVIDIA JETSON14.8 MSFP16 · ISOLATED
YOLO POSE · 8.6 FPSHEAD + WRISTS
OFF AXIS RENDERER4 POINT MAPPING
Build
Debugging the ARM GPU stack offline, including a CUDA and torchvision ABI mismatch, while keeping tracking state and projection geometry honest.
RESULT
A 14.8 ms isolated TensorRT FP16 detection benchmark, about 8.6 FPS live pose, and head plus wrist state driving off axis rendering.
- NOW
- The on device camera, GPU and tracking path is operational. The four point mapping framework is built; physical projector calibration remains.
- NEXT
- Finish physical projector to camera calibration and measure motion to photon latency.
System flow
- 01IMX708 CSI
- 02JETSON GPU RUNTIME
- 03YOLO POSE STATE
- 04SSE + PLANNER
- 05OFF AXIS RENDERER
- 064 POINT MAPPING
- NVIDIA Jetson Orin Nano Super
- CUDA
- cuDNN
- TensorRT FP16
- YOLOv8n Pose
- OpenCV + GStreamer
- Three.js
- SSE