RepACNet: A Lightweight Reparameterized Asymmetric Convolution Network for Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Scene Reconstruction

Background:

• Monocular depth estimation (MDE) is crucial for 2D/3D scene reconstruction, with applications in autonomous driving and robotics.
• Current MDE methods struggle with a trade-off between computational efficiency and accuracy, hindering deployment on resource-constrained devices.
• There is a need for lightweight yet effective MDE models for mobile applications.

Purpose of the Study:

• To develop a novel lightweight network, RepACNet, for efficient and accurate monocular depth estimation.
• To address the limitations of existing MDE methods in terms of computational cost and performance.
• To enable the deployment of MDE on resource-constrained mobile devices.

Main Methods:

• RepACNet integrates a CNN-based architecture with MLP-Mixer components.
• Introduced Reparameterized Token Mixer with Asymmetric Convolution (RepTMAC) for efficient long-range dependency capture with linear complexity.
• Incorporated Squeeze-and-Excitation Consecutive Dilated Convolutions (SECDCs) for multi-scale depth feature extraction using channel attention.

Main Results:

• RepACNet achieves competitive performance on the NYU Depth v2 and KITTI Eigen benchmarks.
• The proposed model maintains significantly fewer parameters compared to state-of-the-art MDE methods.
• RepTMAC enables global feature interaction with minimal computational overhead, outperforming Transformer-based approaches.

Conclusions:

• RepACNet presents an effective lightweight solution for monocular depth estimation.
• The network design successfully balances computational efficiency and estimation accuracy.
• RepACNet is suitable for deployment on resource-constrained mobile devices, advancing applications in computer vision.