Object Detection POP is a native C++ Custom Operators that runs state-of-the-art transformer-based object-detection models fully on GPU inside TouchDesigner. Connect any TOP and the node emits one POP point per detection with bounding box, class, confidence, and an optional stable track id.
WHAT IT DOES
- Runs state-of-the-art models via CUDA or TensorRT for best possible performance with full FP16 and FP32 support.
- Sub-millisecond to a few-millisecond inference depending on model variant and GPU.
- Zero-copy GPU pipeline: reads pixels straight from the source TOP via CUDA, runs preprocess + inference + postprocess on the GPU, no PCIe round-trip.
- Emits one POP point per detection with P (bbox center, ready for instance translate), bbox min/max (x1,y1,x2,y2 normalized), bbox center, bbox size, score, class id, track id, track age and velocity.
- Built-in SORT tracker (per-object Kalman filter + Hungarian assignment) assigns stable per-object ids across cooks, holds ids through brief occlusions, and needs no extra setup.
- Auto-detects the model's input resolution, supports Nano, Small, Base, Medium, Large variants without configuration.
- Built-in filter by class type.
- TensorRT engine caching with a content-addressed cache dir; pre-bake the engine with the included GUI builder to skip the 30-90 s first-cook build entirely on shipped projects.
WHO IT'S FOR
- Interactive installations that react to people, props, or specific object classes.
- Live shows and performances that need real-time scene understanding without a Python sidecar.
- Computer-vision-driven generative pipelines (instance rendering at detection sites, particle emitters, projection mapping onto detected objects).
- Anyone tired of running YOLO in a Python TOP or using external process.
- Anyone willing to do object detection on commercial projects (most YOLO license don't allow it)
Available for download at : https://tdplugins.shop/product/objdetectpop
