Fast CNN Stereo Depth Estimation through Embedded GPU Devices.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Current Convolutional Neural Network (CNN)-based stereo depth estimation models struggle with real-time performance on embedded Graphics Processing Units (GPUs).
• Existing evaluations often overlook model optimization techniques, leaving the true potential of embedded GPUs for this task undetermined.

Purpose of the Study:

• To evaluate the performance of state-of-the-art stereo depth estimation models on embedded GPU devices.
• To assess the impact of optimization methods on model performance.
• To propose an improved architecture for real-time stereo depth estimation on embedded systems.

Main Methods:

• Performance evaluation of two state-of-the-art stereo depth estimation models on three embedded GPU devices (Jetson TX2, Xavier, Nano).
• Comparison of model performance with and without optimization techniques.
• Implementation and evaluation of a U-Net like architecture for cost-volume postprocessing.

Main Results:

• Demonstrated the performance capabilities of embedded GPU devices for stereo depth estimation.
• Achieved real-time inference speeds ranging from 5-32 ms for 1216 × 368 stereo images.
• The proposed U-Net like postprocessing significantly improved runtime speed compared to traditional 3D convolutions.

Conclusions:

• Embedded GPU devices possess significant potential for real-time stereo depth estimation.
• The proposed U-Net based postprocessing method is effective in accelerating stereo depth estimation models.
• This research paves the way for efficient real-time stereo depth estimation in embedded applications.