Mobile Tagging

What is the latest on Zabble Zero™ AI Fullness Predictions?

December 11, 2023

Originally, Zabble Zero™ Mobile Tagging™ provided users with the ability to assess dumpster fullness levels to optimize waste removal and minimize waste. Utilizing an AI classifier based on bin images, we could accurately predict fullness, usually within a 25% range, as mentioned in a previous blog post that described our original (v1) model. 

Since our v1 model, we have developed and deployed v2 and v3 models. Using a more diverse set of images, our latest models accurately predict the fullness of many types of lined bins, carts, and dumpsters. We have met our previous objective of achieving 90%+ accuracy within 25%. With this post, we redefined our metric to mean adjacent precision within 10%: the percentage of total predictions that were within 10% of the ground truth fullness. 

Here are the results of our previous (v2) model when comparing the predicted fullness level to the fullness range provided by human labelers.

Previous model (v2): 64% mean adjacent precision (within 10%) on lined bins, carts and dumpsters

To improve our mean adjacent precision on multiple container types, we significantly expanded the training dataset and fine-tuned our image preprocessing techniques. Classification algorithms train on a large dataset of images that can have preprocessing applied, such as rotating the images during training, to help the model generalize. Adding more image preprocessing led to an enhancement in the mean adjacent precision, such that our predictions are 78% accurate within 10%, across multiple container types.

Updated model (v3): 78% mean adjacent precision (within 10%) on lined bins, carts and dumpsters

Narrowing the Prediction Intervals

Moreover, we previously used a fullness prediction interval of 25%, representing empty, quarter full, half full, quarter empty, full, and overflowing states. However, with the improved model, we could decrease the interval to 10%, providing more accurate predictions. This empowers our customers with a precise understanding of waste generation at their facilities and optimal trash collection timing. The following charts illustrate the improvement over time as we updated our fullness models, by comparing the difference between the predicted fullness and the saved fullness that was reviewed by Zabble Zero users.

Conclusion

By amassing a diverse image dataset of various waste and recycling containers, we refined our AI classifier, enabling better generalization and performance for our customers, while achieving increased accuracy in fullness predictions.

  • Mean adjacent precision (within 25%) is 96%, surpassing our previous objective of 90%+ accuracy within 25%
  • Our new metric of mean adjacent precision (within 10%) improved from 64% to 78%
  • Prediction intervals decreased to 10% for more accurate predictions

Julian Hernandez graduated from Sacramento State with a bachelor’s degree in computer science where he worked on interdisciplinary research using AI to accelerate best recycling practices. He volunteers with DeTrash Berkeley on the weekend to keep his community clean and is currently pursuing a master’s degree at UC Berkeley in the department of civil and environmental engineering.

Our Latest Blog Posts

Technology

Universities Don’t Need Sensors and Robots to Achieve Zero Waste

University sustainability professionals are a dedicated and passionate bunch who go above and beyond the call of duty to help their universities progress towards zero waste. Their knowledge and experience are critical to the success of their waste initiatives. Zero waste teams need to be equipped with data to make informed decisions. The right solution can not only help them deliver on their goals but also be financially beneficial.

Friday, July 19, 2024