Related Experiment Video
Updated: Sep 10, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
635
Human-like monocular depth biases in deep neural networks.
1Communication Science Laboratories, NTT, Inc., Kanagawa, Japan.
Plos Computational Biology
|August 19, 2025
Summary
Human depth perception has systematic distortions. Comparing human and deep neural network (DNN) error patterns reveals shared biases, suggesting efficient strategies for inferring 3D vision from 2D images.
Area of Science:
- Computer Vision
- Human Perception
- Artificial Intelligence
Background:
- Human depth perception from 2D images is complex and not fully understood.
- Deep neural networks (DNNs) show promise in monocular depth estimation.
- Understanding human and artificial depth perception aids in developing better AI models.
Purpose of the Study:
- To compare human and DNN error patterns in monocular depth judgment.
- To develop a framework for analyzing depth estimation biases.
- To investigate the ecological validity of human depth perception strategies.
Main Methods:
- Developed a novel human-annotated dataset of natural indoor scenes.
- Conducted a comprehensive human-DNN comparison for monocular depth judgment.
- Used exponential-affine fitting to decompose depth estimation errors.
Main Results:
- Human depth judgments exhibit systematic biases: depth compression, vertical bias, and affine distortions.
- Accurate DNNs partially replicate human biases, showing similar affine parameters and residual errors.
- DNNs capture metric-level errors but not human-level ordinal depth perception accuracy.
Conclusions:
- Human depth biases may represent efficient strategies for ambiguous 2D images.
- Comparing error patterns offers deeper insights than raw accuracy alone.
- The study provides a framework for aligning AI models with human visual perception.
