Practical overview
When it fits
Use NanoOWL when a Python-based edge-vision workflow needs text-prompted OWL-ViT inference on Jetson Orin, or when detections and classifications need to be arranged in a nested tree.
Who it is for
- Developers building real-time, prompt-driven computer-vision workflows on NVIDIA Jetson Orin.
- Vision practitioners who need nested object detection or text-guided scene classification.
Documented uses
- Detect objects in an image from text labels such as “an owl” or “a glove.”
- Find sub-objects inside detected object regions, such as wings and eyes within an owl detection.
- Classify a scene with text categories and conditionally detect objects within a selected category.
Requirements & limits
The public GitHub repository is licensed under Apache-2.0.
- The documented stack includes Python, PyTorch, torch2trt, NVIDIA TensorRT and Transformers.
- Setup requires building a TensorRT engine for the OWL-ViT image encoder before running the examples.
- The README reports real-time operation on Jetson Orin Nano; its benchmark table includes TBD entries for that device, while listing results for Jetson AGX Orin.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up NanoOWL and run a text-prompted prediction
Install NanoOWL, build the OWL-ViT TensorRT image-encoder engine and run the documented owl-and-glove example.
Before you start
- Python 3
- PyTorch
- torch2trt
- NVIDIA TensorRT
- Transformers
Steps
Install PyTorch, torch2trt, NVIDIA TensorRT and Transformers.
Clone the NanoOWL repository and install the package with the documented develop-mode command.
Create a data directory and run the image-encoder engine build module to create the patch-32 engine.
From the examples directory, run owl_predict.py with the prompts “an owl” and “a glove,” threshold 0.1, and the built engine path.
Expected outcome
The example is documented to save a prediction visualization as data/owl_predict_out.jpg.
Guide limitations
- The repository describes this as a Jetson Orin workflow; it does not establish compatibility with other platforms.
- The recipe covers the basic OWL-ViT prediction example, not the separate tree-prediction pipeline.
Run a nested tree-prediction example
Use the documented tree predictor prompt to detect owls and then identify wings and eyes within detected owl regions.
Before you start
- NanoOWL and its listed setup dependencies
- A built OWL-ViT TensorRT image-encoder engine
- An image available to the example
Steps
Complete the documented NanoOWL setup and build the image-encoder engine.
Change to the examples directory.
Run tree_predict.py with the prompt “an owl [a wing, an eye],” threshold 0.15, and the engine path.
Expected outcome
The example is documented to save a tree-prediction visualization as data/tree_predict_out.jpg.
Guide limitations
- The guide reflects the README’s example prompt and setup; it does not provide performance guarantees for other prompts or workloads.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- NVIDIA-AI-IOT/nanoowl · Official homepage title · retained in its source language