AROS: Affordance Recognition with One-Shot Human Stances

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Robotics

Background:

• Understanding human-object interactions is crucial for intelligent systems.
• Traditional methods require large datasets for training affordance recognition models.

Purpose of the Study:

• To develop a one-shot learning approach for affordance recognition that requires minimal training data.
• To enable systems to predict interaction affordances and generate corresponding human poses in novel 3D scenes.

Main Methods:

• Affordance Recognition with One-Shot Human Stances (AROS) utilizes an explicit representation of human pose and 3D scene interactions.
• The approach is one-shot, meaning it does not require iterative training for new affordance instances.
• Only a few examples of target poses are needed to define interactions.

Main Results:

• AROS was evaluated on three public datasets of scanned real environments.
• Crowdsourced evaluations showed that the one-shot approach was preferred up to 80% of the time over data-intensive baselines.
• The method successfully predicts affordance locations and generates articulated 3D human bodies.

Conclusions:

• AROS offers an efficient and effective solution for one-shot affordance recognition.
• The approach generalizes well to unseen 3D scenes with varying noise levels.
• This work advances the development of more adaptable and data-efficient AI systems for human-robot interaction.