Deep Neural Networks for Accurate Depth Estimation with Latent Space Features.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate 3D scene reconstruction is crucial for human-robot interaction, particularly in indoor environments.
• Monocular depth estimation offers a cost-effective alternative to stereo or LiDAR systems but often struggles with depth boundary precision.
• Existing monocular depth estimation methods face challenges in accurately defining depth boundaries, impacting reconstruction quality.

Purpose of the Study:

• To develop a novel monocular depth estimation framework that enhances the precision of depth maps.
• To improve the accuracy of depth boundaries and local features in monocular depth estimation.
• To provide a more robust solution for 3D scene reconstruction in human-robot interaction.

Main Methods:

• A novel depth estimation framework leveraging latent space features within a deep convolutional neural network.
• A dual encoder-decoder architecture enabling both color-to-depth and depth-to-depth transformations.
• A new loss function combining latent loss and gradient loss to refine depth boundaries.

Main Results:

• The proposed framework sets a new benchmark on the NYU Depth V2 dataset, especially in complex indoor scenarios.
• Demonstrated significant reduction in depth ambiguities and blurring compared to existing methods.
• Achieved enhanced precision in monocular depth map generation.

Conclusions:

• The novel framework effectively improves monocular depth estimation accuracy, particularly at depth boundaries.
• This approach offers a promising solution for precise 3D scene reconstruction in human-robot interaction.
• The method successfully addresses limitations of current monocular depth estimation techniques.