Nested DWT-Based CNN Architecture for Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• 3D image applications like medical diagnosis, navigation, and robotics require accurate depth estimation.
• Deep learning models are widely used for depth prediction from 2D images, but often face challenges with ill-posed, non-linear problems and computational expense.
• Network performance is sensitive to model configuration, loss functions, and training datasets.

Purpose of the Study:

• To develop an efficient and effective deep learning network for depth estimation from 2D images.
• To address the loss of high-frequency information during the downsampling process in conventional encoder-decoder networks.
• To investigate the impact of various architectural components on network performance.

Main Methods:

• Proposed a moderately dense encoder-decoder network named Nested Wavelet-Net (NDWTN).
• Incorporated discrete wavelet decomposition with trainable coefficients (LL, LH, HL, HH) to preserve high-frequency details.
• Studied the influence of activation functions, batch normalization, convolution layers, and skip connections.

Main Results:

• The NDWTN effectively preserves high-frequency information lost in standard downsampling.
• The network demonstrates faster training times compared to existing methods.
• Achieved good performance in depth prediction tasks on the NYU dataset.

Conclusions:

• The proposed Nested Wavelet-Net (NDWTN) offers an efficient approach for depth estimation.
• Discrete wavelet decomposition is a valuable technique for enhancing high-frequency information preservation in deep learning models for depth prediction.
• The network's architecture and training methodology contribute to faster convergence and competitive results.