Ameru: Labeling for a Greener World

Introduction: About Ameru

Ameru is on a mission to accelerate an economically sustainable zero-waste future, and they’re doing so with their smart garbage bins. Each bin automatically sorts waste to be in line with local waste management procedures. The bins, which have scanned and sorted over 130,000 pieces of waste to date, allow companies to reduce the equivalent of 190 cars worth of CO2 from our atmosphere per year, and save companies a combined almost €29000 per year in municipal and office waste disposal costs.

The smart bins are easy to use – the human just needs to dispose of their waste, where it lands on the receptacle. The bin uses an 8MP camera to take a picture of the waste and classifies it on-site using an Nvidia Jenson Nano computer. Finally the categorization is mapped to one of the types of waste disposal – usually general waste, plastic, paper, or glass – and dumps it into the correct bin.

See it in action here!

The Problem

The system itself classifies waste into one of 90 different categories, which are then mapped to the four bins. Training this model takes a lot of data, which needs to be manually labeled for initial training. Once bins go into production, the company needs to audit the predictions that were made on-the-edge to make sure the model continues to perform well for all deployed waste bins. And when the model gets it wrong, they need to add those samples to a new training set for continuous improvement. Processing this amount of image data is challenging, both from a human time and storage perspective, and keeping it organized and easy to use was a key challenge for Ameru.

The Solution: Label Studio

Label Studio had a number of features that Ameru used to enable their product flow.

Flexible Labeling Interface

Ameru was able to combine bounding box image annotation with a more semantic classification schema, where each image is classified as a combination of material, object, and condition. Label Studio allowed them to build out a schema that worked for their particular requirements, regardless of the levels of labels or the complexity of the task. They also loved the flexibility of Label Studio, allowing them to customize the interface with keyboard shortcuts and tools to help their annotators.

Automated Labeling

With Label Studio, Ameru was able to upload their pre-annotated data, labeled by either a bootstrapped model or one of their custom models, and then use humans to review and edit these labels. Leveraging their models allowed for a faster time to high quality labeled data, and allowed them to invest their human hours into the highest priority parts of their project and data evaluation.

Organized Project Structure

Ameru loved the organization that Label Studio provided, allowing each of their customers to be kept as a separate project in Label Studio. Keeping all their projects in one place, with the same structure, enabled them to stay organized and ensure that they are providing best in class service to all of their clients.

Data On Data

The data that Label Studio provides also enabled Ameru to set up their customer-side reporting about waste management, usage, and optimization. By doing some backend processing on top of Label Studio’s data, Ameru is able to generate reports that inform the customer how much they’re saving, both environmentally and economically, by investing in smart bins.

They also are able to provide insights into how best to cut back on the waste that they’re creating by suggesting greener alternatives to frequently disposed-of items. They break this information into a number of different charts, providing clear pathways to a greener future.

What’s Next For Ameru?

Part of what Ameru loves about Label Studio is the ability to scale. In their quest to enable an economically sustainable zero waste future, they hope to continue increasing the number of smart bins that are in production. Label Studio will allow them to scale as their business grows, helping them to improve their models and the environment as a whole.