🤖 AI

Efficient Waste Sorting for Circular Economy: A Confidence-guided comparison between One-Vs-All and One-Vs-Rest Classification Strategies with Human-in-the-Loop for Automated Waste Sorting

This paper evaluates One-Vs-All and One-Vs-Rest classification strategies for automated waste sorting in Germany, utilizing a confidence-guided human-in-the-loop approach to balance misclassification reduction with human annotation effort while accommodating municipal-specific sorting schemes.

Original authors: Mohammed Fahad Ali, Dominique Briechle, Marit Briechle-Mathiszig, Tobias Geger, Andreas Rausch

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammed Fahad Ali, Dominique Briechle, Marit Briechle-Mathiszig, Tobias Geger, Andreas Rausch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Sorting Trash with a Smart Assistant

Imagine you live in a town where the rules for throwing away trash are tricky. In one neighborhood, a pizza box goes in the "Paper" bin; in the next town over, it belongs in the "General Waste" bin. This confusion leads to a lot of trash being thrown in the wrong place, which hurts our planet's ability to recycle things (what the paper calls the "Circular Economy").

The authors of this paper wanted to build a smart mobile app that acts like a helpful assistant. You take a picture of your trash, and the app tells you exactly which bin to use. To make this work, they had to teach a computer (Artificial Intelligence) how to recognize different types of waste.

The Core Question: How Should the Computer "Think"?

The main problem the paper tackles is: What is the best way to teach the computer to sort trash?

They compared two different "teaching styles" (classification strategies):

The "All-in-One" Coach (One-vs-All / OvA):
- The Analogy: Imagine a single teacher standing in front of a classroom of six students (representing six types of trash: Paper, Plastic, Glass, etc.). The teacher looks at a piece of trash and asks, "Which one of you does this belong to?" The teacher picks the student with the highest confidence.
- How it works: The computer learns one big rule to distinguish all six types at once.
The "Specialist" Team (One-vs-Rest / OvR):
- The Analogy: Imagine a team of six different experts. Each expert only knows one thing perfectly.
  - Expert A only knows: "Is this Paper or NOT Paper?"
  - Expert B only knows: "Is this Plastic or NOT Plastic?"
  - And so on.
- How it works: The computer asks all six experts. If the "Paper" expert says "Yes!" and everyone else says "No," then it's paper. If two experts shout "Yes!" at the same time, the computer gets confused.

The Experiment: Goslar's Trash

The researchers didn't just guess; they built a real test using data from the city of Goslar, Germany.

They collected over 22,000 photos of trash (like bottles, cans, food scraps, and glass).
They made sure the photos looked like real life (some were dirty, some were in shadows, some were on the street) so the AI wouldn't just memorize clean pictures.
They trained both the "All-in-One" coach and the "Specialist" team on the same photos and tested them on new, unseen photos.

The Results: Who Won?

1. Overall Accuracy (Who got the most right?)

The Winner: The "All-in-One" Coach (OvA) was slightly better. It made fewer mistakes overall.
The Runner-up: The "Specialist" Team (OvR) made a few more mistakes (about 5 more errors out of 2,000 tests).

2. The "Confidence" Test (Who knows when they are unsure?)
This is the most interesting part of the paper. The researchers wanted to know: If the computer is unsure, can we tell it to ask a human for help?

The "All-in-One" Coach: When this coach is wrong, it is often very confident about being wrong. It's like a student who is 100% sure they are right, but they are actually wrong. Because it's so confident, the system doesn't flag it for human review, and the mistake slips through.
The "Specialist" Team: This team is much better at admitting when it's confused.
- Sometimes, two experts will both shout "Yes!" (e.g., "It's Plastic!" and "It's Glass!"). This conflict signals high uncertainty.
- Sometimes, no expert is sure (all say "Maybe").
- The Magic: The researchers found that if they asked a human to only check the photos where the "Specialist" team was confused (specifically when experts disagreed or were unsure), they could catch over 50% of all the mistakes while only asking the human to look at less than 5% of the total photos.

Why Does This Matter?

The paper concludes that while the "All-in-One" coach is slightly more accurate overall, the "Specialist" team is much better at knowing its limits.

Flexibility: If a town changes its rules (e.g., "Now we recycle pizza boxes!"), it's easy to swap out one "Specialist" expert without retraining the whole system.
Human-in-the-Loop: The "Specialist" approach is perfect for a system where a human reviews the tricky cases. It saves human time because it only flags the trash that is actually confusing, rather than asking humans to check everything.

Summary

The paper shows that for a smart trash-sorting app:

If you just want the highest possible score, use the Single Coach (OvA).
If you want a system that is flexible and knows when to ask for human help to fix its mistakes, use the Team of Specialists (OvR).

This balance is crucial for making the "Circular Economy" work, where resources are reused efficiently, and human effort is used wisely to fix the AI's blind spots.

Technical Summary: Efficient Waste Sorting for Circular Economy

Problem Statement
The transition to a Circular Economy is hindered by the complexity of waste disposal regulations, which vary significantly across European municipalities. In Germany, for instance, sorting requirements differ not only between federal states but also between individual districts. This heterogeneity makes it difficult for residents to dispose of waste correctly, leading to high rates of contamination (e.g., up to 30% of plastic waste in Germany is improperly disposed of) and increased treatment costs. While AI-based mobile applications offer a potential solution by guiding residents, existing systems often lack the configurability required to adapt to specific local sorting schemes. Furthermore, standard multi-class classification approaches struggle with modularity; adapting a model to new regulations typically requires full retraining. Additionally, there is a need to balance predictive performance with human annotation effort, specifically by identifying uncertain samples for human review without excessive manual labor.

Methodology
The authors evaluated and compared two prominent classification strategies: One-vs-All (OvA) and One-vs-Rest (OvR), utilizing a dataset aligned with the waste sorting scheme of Goslar, Germany.

Dataset Construction: The study utilized a six-class waste management dataset constructed by merging and mapping images from the TACO and TrashNet datasets to a local framework. The final categories included Organic Waste, Paper Waste, Plastic and Metal Waste, Recycling Center, Glass, and Residual Waste. The dataset was stratified into 18,515 training, 2,000 validation, and 2,000 test images.
OvA Strategy: A single multi-class classifier was built using a pre-trained InceptionV3 backbone (ImageNet) with custom Global Average Pooling, Fully Connected (FC), and Dropout layers. The output layer utilized a Softmax activation function to generate a normalized probability distribution across the six classes.
OvR Strategy: Six independent binary classifiers were trained, one for each target class. Each classifier used the same InceptionV3 backbone and architecture but employed a Sigmoid activation function for its single output neuron. Positive samples for a specific classifier were instances of the target class, while negative samples comprised all other classes.
Confidence-Guided Analysis: To address the trade-off between misclassification rates and human annotation effort, the authors applied varying confidence thresholds. Test samples were categorized into four groups based on prediction probabilities:
1. High-Confidence (Group 1): One class with $p \ge 0.95$ .
2. Single-Vote (Group 2): One class with $0.50 < p < 0.95$ .
3. Multi-Vote (Group 3): Multiple classes with $0.50 < p < 0.95$ (exclusive to OvR).
4. No-Vote (Group 4): All classes with $p \le 0.50$ .

Key Results

Overall Accuracy: The OvA strategy achieved marginally higher overall accuracy, misclassifying 49 images compared to 54 for the OvR strategy.
Error Patterns: Only 29 images were commonly misclassified by both strategies, indicating distinct error patterns for each approach.
Uncertainty Identification:
- OvA: The strategy struggled to identify uncertain samples. Group 3 contained no samples (as probabilities sum to 1), and Group 4 contained only 2 samples (both misclassified). Consequently, targeting Groups 3 and 4 captured only ~4% of misclassifications. High-confidence errors (Group 1) were difficult to detect, with 27 misclassified samples "lost" due to high predicted probabilities.
- OvR: This strategy demonstrated superior capability in flagging uncertain samples. Group 3 (Multi-Vote) and Group 4 (No-Vote) contained a significant concentration of errors. Specifically, focusing on Groups 3 and 4 allowed for the identification of over 50% of misclassifications while requiring human annotation of less than 5% of the total data. Focusing solely on Group 4 identified over 35% of misclassifications with less than 2.5% annotation effort.

Significance and Claims
The paper claims that while the OvA strategy offers slightly better raw accuracy for multi-class waste sorting, the OvR strategy is superior for iterative model refinement and adaptability. The significance of the work lies in demonstrating that OvR:

Enhances Modularity: It allows for the selective addition or removal of waste categories (e.g., to comply with new municipal regulations) without retraining the entire system.
Optimizes Human-in-the-Loop Efforts: It provides a more effective mechanism for identifying uncertain samples for human review, significantly reducing the annotation effort required to correct a large portion of errors.
Supports Circular Economy Goals: By offering a configurable framework that can adapt to decentralized local regulations, the OvR approach supports the deployment of scalable, user-friendly tools for waste sorting.

The authors conclude that future work should explore ensemble frameworks combining both strategies to leverage their complementary strengths and investigate adaptive thresholds as hyperparameters.