DDCNet: Deep Dilated Convolutional Neural Network for Dense Prediction

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Dense pixel matching, including optical flow and disparity estimation, presents significant challenges in computer vision.
• Deep learning methods have shown recent success in addressing these dense estimation tasks.
• Larger effective receptive fields (ERFs) and high spatial feature resolution are crucial for accurate, high-resolution dense predictions.

Purpose of the Study:

• To present a systematic approach for designing network architectures that enhance the effective receptive field (ERF) while preserving high spatial feature resolution.
• To develop compact deep learning models for dense pixel matching tasks.

Main Methods:

• Utilized dilated convolutional layers to systematically increase the effective receptive field (ERF).
• Aggressively increased dilation rates in deeper network layers to achieve a larger ERF efficiently.
• Employed the optical flow estimation problem as a benchmark to validate the network design strategy.

Main Results:

• The proposed network architectures achieve a significantly larger ERF with a reduced number of trainable parameters.
• Compact networks demonstrate comparable performance to existing lightweight models on benchmark datasets.
• Validated performance on challenging benchmarks including Sintel, KITTI, and Middlebury datasets.

Conclusions:

• The developed network design strategy effectively balances ERF expansion and spatial resolution maintenance.
• The proposed compact networks offer a promising solution for efficient and high-performance dense estimation in computer vision.
• The approach provides a viable alternative for lightweight models in optical flow and disparity estimation.