Performance specifications of the AI recognition

The performance specifications below apply to the analysis of continuous waste streams and waste samples

Introduction

Greyparrot’s waste monitoring system uses AI-powered computer vision to detect, identify and classify waste items moving through sorting facilities in near real-time. It enables data-driven decisions by sending waste stream composition analysis to a live dashboard, or to third-party machinery and software through various APIs.

The platform is built on – and continuously trained by – millions of data points from sorting facilities around the world. Greyparrot’s AI model characterises waste items based on visual features such as shape, size, colour and light reflection. It accounts for the conveyor belt and other environmental variables in the process. The classification library includes 49 detailed classes consisting of different waste superclasses including plastic, fibre and metal. We add more granularity and new waste types on a regular basis.

Greyparrot’s highly adaptable system is capable of accurately analysing a wide range of waste streams. A fine-tuning process takes place at every deployment to ensure the AI model is adjusted to individual belts and facilities.

The performance specifications below apply to the analysis of continuous waste streams and waste samples. The performance for automated sorting integration applications differ by their nature and are not evaluated with the same metrics.

Performance Specifications 

Waste stream composition analysis (Mass):

• Error <2%
• The composition shown on the dashboard deviates by less than 2% from the reality. This applies to every waste category. 

Example:

Waste stream throughput (Mass):

• Error <5%
• The mass shown by our system deviates by less than 5% from the reality. This applies to every waste category. 

Example:

Parameters

The indicated performance is valid for:

• Categories that represent at least 1 kg and at least 2% of the waste stream
• Continuous monitoring of > 1 day
• Sampling analysis of 25kg to 50kg
• Mono-layer waste streams - at least 50% of each object needs to be visible for 3D items, and at least 75% of each object for 2D items. Multi layered waste stream and higher overlaps might negatively impact the accuracy.
• Testing based on the grouping presented below, and existing Greyparrot waste classes as defined by Greyparrot (definition available).

The model will classify as fine any item too small to be recognised as another object. This generally includes objects smaller than 30mm.

Performance requirements

Achieving the indicated performance requires that the customer provide enough data from their own mass estimation to establish a benchmark for the model optimisation. This can be at a weekly, daily, hourly or sample level.  

In other words, at least one of the below:

• At least 3 samples
• At least 4 weeks (if based on weeks)
• At least 15 operating days (if based on days)
• At least 48 operating hours (if based on hours)

The time it takes to reach the indicated performance will depend on the time necessary for the customer to provide the data, as well as the quantity and quality of the data provided.

The technical requirements and guidelines (provided by us) need to continuously be met and followed for these performances to be valid.

Waste Classes - Category