AI-Powered Waste Segregation: Accelerating Annotations with SAM and CSRSF

Back to Case Studies
Waste Segregation AI

Executive Summary

HaiData partnered with a leading cleantech startup to deliver 1.2 million high-precision waste segmentation annotations for their AI-enabled waste segregation system. By combining the Segment Anything Model (SAM) for initial automation with skilled human annotation and implementing our innovative Cognitive Shift Role-Swapping Framework (CSRSF), we achieved exceptional accuracy while maintaining annotator productivity and well-being. The project demonstrates how intelligent workflow design and advanced AI tools can overcome the challenges of large-scale, complex annotation tasks.

Client Overview

  • Industry: Cleantech / Environmental Technology
  • Focus: AI-enabled waste segregation systems for automated sorting
  • Objective: Develop computer vision models to accurately identify and classify various waste materials on conveyor belts in real-time

Project Overview

Annotation Type Instance Segmentation (Pixel-level masks)
Total Annotations 1.2 million waste object instances
Waste Categories Multiple categories including plastics, metals, paper, organic waste, glass, textiles, and mixed materials
Project Duration 6 months

The Challenge

The project presented several significant technical and operational challenges:

1. Overcrowded and Overlapping Objects

Many images captured from the conveyor belt contained densely packed waste materials with significant overlap. Manual segmentation of individual items in these crowded scenes was extremely time-consuming and mentally taxing, requiring annotators to carefully trace complex boundaries between overlapping objects.

2. Visual Similarity and Ambiguity

Different types of waste materials often exhibited similar visual characteristics, making classification challenging. For example, distinguishing between different types of plastics or identifying contaminated materials required careful attention and domain expertise.

3. Scale and Monotony

With 1.2 million annotations required, the sheer volume of repetitive work posed risks of annotator fatigue, decreased accuracy over time, and potential burnout. Traditional annotation approaches would struggle to maintain quality and efficiency at this scale.

4. Accuracy Requirements

The client's AI system required high-precision annotations to achieve reliable waste sorting. Even small errors in segmentation boundaries could impact model performance and downstream sorting accuracy.

The Solution

HaiData developed a comprehensive two-level approach that combined cutting-edge AI technology with innovative workforce management:

Level 1: SAM-Powered Automated Segmentation

We leveraged Meta's Segment Anything Model (SAM), a foundation model for image segmentation, to generate initial segmentation masks automatically. SAM's zero-shot capabilities allowed it to segment diverse waste objects without requiring task-specific training, providing:

  • Rapid initial segmentation: SAM processed thousands of images quickly, creating baseline masks for all visible objects
  • Handling of complex scenes: The model effectively identified object boundaries even in crowded, overlapping scenarios
  • Acceleration of workflow: Automated pre-annotation reduced the manual workload by approximately 60-70%, allowing annotators to focus on refinement rather than starting from scratch
Level 2: Skilled Human Annotation and Quality Assurance

Our trained annotators reviewed and corrected SAM-generated masks, ensuring:

  • Boundary precision: Fine-tuning segmentation boundaries for pixel-perfect accuracy
  • Classification accuracy: Correct labeling of waste categories based on material type and characteristics
  • Handling edge cases: Addressing ambiguous objects, partial occlusions, and contaminated materials
  • Quality validation: Multi-tier review process to maintain consistency across the entire dataset

Innovation: Cognitive Shift Role-Swapping Framework (CSRSF)

To address the challenge of monotonous, large-scale annotation work, HaiData implemented our proprietary CSRSF methodology. This productivity and well-being oriented workflow model fundamentally reimagines how teams approach repetitive tasks.

How CSRSF Works
  • Task Decomposition: Large annotation tasks are broken down into smaller, manageable sub-tasks (e.g., T1_a: plastic bottles, T1_b: plastic bags, T1_c: rigid plastics)
  • Group Allocation: Team members are organized into small collaborative groups (2-3 people per task), with each member starting on a different sub-task
  • Structured Role Swapping: At defined intervals (e.g., post-lunch, after completing a batch, or at milestone points), team members systematically swap roles within their group
  • Collaborative Completion: Through coordinated rotation, each group collectively completes all sub-tasks with fresh perspectives and natural peer review
Key Benefits of CSRSF
  • Reduced Mental Fatigue: Introducing variety through role rotation prevents the cognitive exhaustion associated with repetitive work
  • Cognitive Stimulation: Switching between different waste categories or annotation aspects activates different mental processes, keeping annotators engaged and alert
  • Natural Peer Review: As team members rotate through sub-tasks, they review and refine each other's work, creating built-in quality checks
  • Enhanced Collaboration: Shared ownership of tasks fosters teamwork, knowledge sharing, and collective problem-solving
  • Improved Output Quality: Fresh eyes on each sub-task catch errors and inconsistencies that might be missed in traditional workflows
  • Higher Job Satisfaction: Variety, collaboration, and reduced monotony lead to better morale and lower turnover
Why CSRSF Works: The Science Behind It

CSRSF leverages several psychological and organizational principles:

  • Task Segmentation: Breaking work into smaller units makes progress more visible and achievable, boosting motivation
  • Micro-Collaboration: Small team structures enable effective communication and mutual support without coordination overhead
  • Cognitive Diversity: Different perspectives and approaches to the same task improve overall quality through complementary strengths

Results and Impact

Metric Achievement
Annotations Delivered 1.2 million high-quality segmentation masks
Annotation Accuracy 98.5% accuracy maintained throughout project
Productivity Gain 60-70% reduction in annotation time vs. fully manual approach
Team Satisfaction 95% positive feedback on CSRSF methodology
Project Completion Delivered on time within 6-month timeline

Cloud Storage Integration

As the data was critically important to the client's operations, they hosted their dataset on secure cloud storage (S3 bucket). We seamlessly synced the dataset directly from the S3 bucket to our annotation platform, ensuring data integrity and security throughout the project lifecycle.

  • Direct S3 Integration: Our platform natively supports direct synchronization with AWS S3 buckets
  • Data Remains Client-Controlled: The dataset stayed in the client's own cloud storage, maintaining their control and ownership
  • Seamless Workflow: Annotators accessed images directly from S3, eliminating the need for manual data transfers
  • Secure Access: All data access was managed through secure credentials and IAM policies
  • Efficient Bandwidth Usage: Smart caching and progressive loading optimized network bandwidth while maintaining annotation speed

Cost-Effective Infrastructure

We used locally hosted annotation platform with GPU support on our servers, that significantly reduced the overall project cost. HaiData doesn't charge customers for the platforms. This is an added advantage that sets us apart from traditional annotation service providers.

  • No Platform Fees: Clients pay only for the annotation work, not for platform access or usage
  • GPU-Accelerated Processing: Our locally hosted infrastructure with GPU support enabled faster SAM model processing
  • Reduced Overall Costs: By eliminating platform licensing fees and leveraging efficient infrastructure, we delivered significant cost savings to the client
  • Data Security: Local hosting ensures complete data privacy and security throughout the annotation process
For more information on how HaiData can help with your AI data annotation needs, please write to info@haidata.ai