nanoowl

Why we included this project

NanoOWL brings open-vocabulary object detection to embedded NVIDIA hardware by compiling OWL-ViT's image encoder into a TensorRT engine, so you get real-time inference on Jetson Orin Nano and AGX Orin. It also chains OWL-ViT with CLIP in a tree detection pipeline, which means you can phrase nested queries like "an owl [a wing, an eye]" and have the model find the bird, then locate and classify parts inside each detection. That text-driven approach is handy for any project that wants to detect arbitrary objects without training a custom model, and the Python API plus worked examples keep the learning curve short. If you also need masks, NanoOWL pairs with NanoSAM for zero-shot instance segmentation.