A comparison between human and machine labelling of image regions

Area of Science:

• Computer Vision
• Cognitive Science
• Artificial Intelligence

Background:

• A prior vision system segmented outdoor scenes into regions for object classification using an artificial neural network (ANN).
• The ANN was trained to label image regions with one of eleven object types based on feature descriptions.

Purpose of the Study:

• To determine the importance of individual visual features for object classification performance in both human and machine vision.
• To assess whether an ANN can serve as a model for human image region labeling.

Main Methods:

• Human subjects and an ANN were trained on an image region labeling task.
• Experiments involved presenting intact and degraded stimuli (image regions) to both human subjects and the ANN.
• A novel method corrupted individual features in the ANN's input to simulate information loss.

Main Results:

• Certain visual features were found to be highly influential for overall labeling performance in both humans and the ANN.
• The results from the ANN experiment showed broad similarities to those observed in human subjects.
• The study demonstrated that the ANN can effectively model human image region labeling.

Conclusions:

• Specific features play a critical role in mediating object recognition performance across both human and artificial vision systems.
• The computational and psychophysical methodology presented offers a valuable tool for future research in vision science.
• Artificial neural networks show promise as models for understanding human visual perception and object recognition.