Exploiting Depth From Single Monocular Images for Object Detection and Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Measured depth data, often from sensors like Microsoft Kinect, improves computer vision tasks.
• Most available images lack depth information, limiting their utility in depth-aware applications.
• Existing methods rely on direct depth measurement, which is not universally accessible.

Purpose of the Study:

• To demonstrate that estimated depth can augment RGB images to improve computer vision task performance.
• To develop and evaluate methods for integrating estimated depth features with RGB data.
• To propose a novel RGB-D semantic segmentation approach using multi-task learning.

Main Methods:

• Learned a deep depth estimation model from monocular RGB images.
• Extracted deep depth features from estimated depth maps.
• Combined RGB and estimated depth features for object detection and semantic segmentation.
• Developed an RGB-D semantic segmentation method using a multi-task learning framework (semantic prediction and depth regression).

Main Results:

• Augmenting RGB data with estimated depth significantly improved object detection accuracy.
• Incorporating estimated depth information led to remarkable improvements in semantic segmentation performance.
• The proposed multi-task learning approach for RGB-D semantic segmentation showed enhanced results.

Conclusions:

• Estimated depth from monocular images is a viable and effective substitute for measured depth in computer vision.
• Integrating estimated depth features alongside RGB data offers substantial performance gains for object detection and semantic segmentation.
• The developed methods provide a practical way to leverage depth information from standard images, broadening the applicability of depth-aware computer vision techniques.