Digital Transformation

How accurate is Zabble's AI in predicting bin fullness ?

How well does Zabble Zero™ AI perform?


Zabble Zero™ Mobile Tagging makes waste level and contaminant monitoring easy using just a mobile device. Zabble Zero™ AI also makes it faster and learns over time: When a user snaps a photo of a bin using a phone or a tablet, Zabble Zero™ AI suggests fullness levels and contamination items, and automatically alerts supervisors of any issues. 

In this post, we take a deeper look into the performance of the suggested fullness algorithm. How often does a user need to modify the AI suggestion? Does it really work?

For this evaluation, we focused on dumpster tags from April and May. We looked at two different metrics:

  1. The difference between the suggested fullness and what the user ultimately saved as the fullness
  2. The difference between the suggested fullness and a human estimation from looking at photos

Difference between AI-predicted fullness % and user-selected fullness %

Zabble AI predicted vs. user-selected bin fullness

Zabble Zero™ customers used the app’s AI to predict the fullness of dumpsters 735 times over two months. In 719 cases, users accepted the suggested fullness without making any changes. Only in a handful cases did they change the suggested fullness, and those changes were typically within 25%. (Zabble Zero™ AI predicts fullness in 25% increments.) This evaluation method suggests a 98% accuracy.

As users may not always review the suggested fullness, we conducted a photo evaluation as well. We took a random sample of 30 dumpster photos and, without looking at the suggested fullness results, entered our own estimation of the fullness %. And repeated this experiment 10 times.


Difference between AI-predicted fullness % and human-estimated fullness % based on photo

Zabble AI predicted vs. human-estimated bin fullness

In this one sample, in 68% of cases, the human estimation matched the AI suggestion exactly. In 89% of cases, the difference was close, or within 25%. (Note, estimating fullness by looking at a photo is not always accurate, as photo usability can vary.) 

The average difference was 0%, meaning that differences will even out over many tags. 

We repeated this evaluation multiple times, examining 10 samples of 30 entries each, to see if our results were consistent across different sets of photos.

Photo evaluation results over 10 samples

Zabble AI predicted vs. human-estimated bin fullness across 10 sets of 30 samples

Our photo evaluation results proved to be fairly consistent. Across the samples, the AI suggestions and human estimations matched 63% of the time, and were close 94% of the time. These are great results that show how Zabble Zero™ AI can help make zero waste campaign tracking fast and easy. 

Concluding Results

  • A large sample size of 735 reported dumpster occurrences showed a 98% accuracy in Zabble Zero™ AI receptacle fullness recommendations
  • 10 experiments of 300 photo evaluations to determine AI-predicted fullness % versus human-estimated fullness % resulted in 94% of estimations being within the 25% tiered Zabble Zero™ AI receptacle fullness recommendation

To learn more about Zabble Zero™ AI, please contact the Zabble team.

Our Latest Blog Posts

Zero Waste

Zabble Wants to Play with Your Garbage

This blog post discusses Zabble's team visits to key partners, PSSI, Stanford and UCSF, to observe a day in their lives and how they use Zabble's products for zero waste initiatives. At PSSI, Zabble's team rode along with garbage and recycling truck drivers, observing challenges like route changes and contamination. They learned about the importance of real-time communication and notifications for addressing these issues. At UCSF, the team witnessed manual waste sorting and the need for technology that works offline. They also saw how Zabble's "Alerts" feature helps communicate contamination and hazardous items. The blog concludes with the value of these visits and the vision for Zabble 3.0, which aims to provide more granular insights and proactive engagement for achieving zero waste.

Wednesday, April 23, 2025

Artificial Intelligence

How Zabble's Object Detection Model Stacks Up Against Industry Benchmarks

A summary of Zabble's AI object detection models development process and how we use industry metrics and benchmarks to develop, test and evaluate our AI systems.

Wednesday, April 2, 2025

Artificial Intelligence

A Beginner’s Guide to Object Detection AI

Read about the basics of object detection AI, challenges to developing a methodology and metrics and what it means for contamination detection and identification at the bin.

Monday, March 31, 2025