Monocular Depth Estimation with Joint Attention Feature Distillation and Wavelet-Based Loss Function.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Monocular depth estimation is vital for 3D vision but faces challenges due to inherent ambiguity.
• Existing methods often yield unsatisfactory results because of the ill-posed nature of the problem.

Purpose of the Study:

• To propose a novel deep convolutional neural network for accurate monocular depth estimation.
• To improve feature discrimination and enhance structural detail recovery in depth maps.

Main Methods:

• A deep convolutional neural network incorporating joint attention feature distillation and a wavelet-based loss function.
• Progressive feature distillation and refinement strategy for hierarchical feature extraction.
• Joint attention mechanism for feature aggregation.

Main Results:

• The proposed method significantly improves feature modulation discrimination.
• The wavelet-based loss function enhances the recovery of structural details in depth maps.
• Experimental results demonstrate superior performance on indoor and outdoor benchmark datasets compared to state-of-the-art methods.

Conclusions:

• The novel network effectively addresses monocular depth estimation challenges.
• Joint attention and wavelet loss are key innovations for superior depth recovery.
• The method shows significant advancements in 3D vision applications.