Pseudo-depth-based deep neural network model for object detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current machine learning models for computer vision primarily use three-channel RGB color features.
• Optical images inherently lack explicit third-dimensional spatial information, limiting recognition performance.
• Exploiting spatial features is crucial for advancing object detection capabilities.

Purpose of the Study:

• To enhance object detection performance by integrating pseudo-depth features with RGB data.
• To develop a novel detection scheme that leverages four independent features without requiring additional hardware.
• To improve the extraction of spatial features in machine learning models.

Main Methods:

• Utilized a monocular depth estimation model to generate pseudo-depth features from optical images.
• Developed a fused Depth-RGB feature representation by combining pseudo-depth and RGB data.
• Integrated the fused features into neural network models for object detection training and inference.

Main Results:

• Achieved a 3.8 percentage point improvement in the mean Average Precision (mAP) on the M[Formula: see text]FD dataset.
• Demonstrated an 8.0 percentage point increase in mAP on the COCO dataset.
• The proposed scheme showed significant enhancements in detection accuracy.

Conclusions:

• The proposed Depth-RGB feature fusion method effectively enhances spatial feature extraction for object detection.
• This approach offers a hardware-agnostic solution to improve machine learning model performance.
• The scheme is versatile and can be readily integrated into existing machine learning models to boost detection capabilities.