YOLO-based vision classification software that captures and recognizes the material a consumer drops in (glass, PET, metal) while it's moving on the conveyor. Trained on 100k+ real field images; classification time around 100ms. Runs on the manufacturer's existing processing platform with low resource consumption.
Our software recognizes the material a consumer drops into the manufacturer's RVM while it's moving on the conveyor. The classification result (glass / PET / metal + confidence score) is delivered to the manufacturer's software, where the routing-to-the-right-compactor or accept/reject decision is made.
Runs on top of Ultralytics YOLOv8 (PyTorch). Despite the broad training set, the model is small because it's optimized; it performs real-time classification with low resource consumption.
Image acquisition with industrial camera supported by transillumination (back lighting). Reliably distinguishes transparent glass and colored plastic; resilient to shadow and lighting variability.
Our model runs on Ultralytics YOLOv8 (PyTorch) and classifies a moving bottle on the conveyor in approximately 100ms. Despite the broad training set, the model is small because it's optimized; it consumes few resources on the manufacturer's existing vision platform. No cloud dependency.
Our model has been trained on 100,000+ real field images and continues to grow. New versions are distributed to manufacturers OTA (over-the-air); no field visit required.
All classification happens on the device. The machine keeps running even if the network goes down; consumer data never leaves the device.
Our model is trained not on lab images but on 100,000+ images of materials actually dropped into RVMs by real consumers. The training set keeps growing with each release.
High recognition rate on items with peeled labels, partial deformation, or light contamination; safe classification of transparent containers via transillumination.
New model versions are distributed remotely to the manufacturer's fleet. No field visit, USB update, or restart required.
For machines that handle three material types simultaneously, the routing-to-the-right-compactor decision.
Cross-validates the 360° barcode reader's result with the visual classification; eliminates erroneous or fraudulent acceptances.
Recognizes objects that should not be accepted (e.g., old/shattered bottles, foreign waste) and protects the machine.
Custom model variants tuned to local packaging diversity in Germany, Netherlands, UK, Ireland and Bulgaria.
Runs on top of Ultralytics YOLOv8 (PyTorch). Integrated into the manufacturer's existing vision processing platform; performs real-time classification thanks to its low resource consumption.
Delivered to the manufacturer as licensed software (model + inference runtime). Standard call interface: send an image, receive class + confidence score + timestamp. No cloud dependency.
New model versions distributed OTA on a per-fleet basis. The manufacturer rolls our updates onto its machines; no field visit required.
Recognition accuracy rates, inference times, model versioning, OTA architecture, message protocol details and resource consumption profile are shared with NDA-signed manufacturers via a product-specific datasheet.
Describe your need, and our technical team will start the NDA process and share a product-specific datasheet and integration document with you.
Tell us your component need →