![]()
Windows · Current release 1.9.0
InSiSo Model Bench¶
A Windows tool for labelling images, training a detection model and exporting it. No coding and no cloud.
Current release 1.9.0 · 6 Oct 2026
What’s in 1.9.0
- One installer uses the NVIDIA GPU when present, otherwise the CPU. The training log says which.
- Projects and folders for classes, images, review and model versions.
- Labelling with box, polygon and rotated box, plus trackpad navigation and shortcuts.
- ONNX (Standard); Hailo and NVIDIA GPU (Pro).
Install and license · Train and export
What it does¶
Bounding boxes, rotated boxes and polygons on your own images.
Object detection on NVIDIA GPU if present, otherwise CPU. Nothing to configure.
Let the model suggest labels, then review them.
ONNX Runtime (CPU), Hailo, or NVIDIA GPU.
Import frames captured on the machine, correct them, retrain.
Millimetres from 2D images or 3D ToF clouds; CSV/Excel (Pro).
The model is class-agnostic: it learns whatever classes you define (cartons, scratches, screws, leaves). Model Bench runs fully offline; images, labels and models stay on the engineering PC.

How the products work together¶
Labelling, training, export, measurement
Runs the exported model on live cameras
Holds the Model Bench licence
An exported model carries its class names and pre-processing settings inside the file, so the runtime app needs no extra configuration to read it.
Where Model Bench fits, and where it does not¶
- It is object detection. It finds and locates things a person can label with a box or polygon. It needs labelled examples of each defect; it is not unsupervised anomaly detection.
- No reading in Model Bench. Model Bench does no OCR or barcode reading. On the line, InSiSo Visual Detector reads barcodes; see Parcel dimensioning and profiling.
- Package-scale measurement. Millimetre-level dimensions of objects such as parcels, not micron-level gauging.
- Not a safety function. Detections are not certified for personnel protection; use rated safety devices for that.