Surpassing Humans and Computers with JellyBean: Crowd-Vision-Hybrid Counting Algorithms

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Image Processing

Background:

• Object counting is crucial for various applications but faces challenges with traditional methods.
• Supervised computer vision requires extensive labeled data and struggles with complex images.
• Crowdsourcing alone can be inaccurate, especially for images with numerous objects.

Purpose of the Study:

• To develop a cost-effective and accurate object counting algorithm.
• To combine the strengths of crowdsourcing and computer vision for improved performance.
• To introduce the JellyBean suite of algorithms for robust object counting.

Main Methods:

• Utilizing a judicious decomposition of images to enhance counting accuracy.
• Developing algorithms that are theoretically optimal or near-optimal in human query efficiency.
• Implementing both stand-alone and hybrid modes for flexibility with computer vision integration.

Main Results:

• Achieving high accuracy in object counting, even on challenging images.
• Demonstrating cost-effectiveness compared to existing methods.
• Validating the theoretical optimality and practical performance of the algorithms.

Conclusions:

• The JellyBean suite offers a superior solution for object counting by integrating crowds and computer vision.
• The hybrid approach provides flexibility and enhanced accuracy.
• This method significantly outperforms individual workers or computer vision algorithms in complex scenarios.