Semantics-enhanced supervised deep autoencoder for depth image-based 3D model retrieval.

Area of Science:

• Computer Vision
• Pattern Recognition
• 3D Shape Analysis

Background:

• Depth sensors like Kinect provide rich data for 3D applications.
• 3D model retrieval is a growing area in computer vision.
• Cross-domain retrieval from depth images faces challenges like occlusion, noise, and view variability.

Purpose of the Study:

• To propose a novel supervised deep autoencoder and semantic modeling approach for depth image-based 3D model retrieval.
• To address the challenges of incompleteness and ambiguity in depth data for cross-domain retrieval.
• To develop a single deep network capable of cross-domain retrieval using real and synthetic data.

Main Methods:

• A supervised deep autoencoder is developed for robust feature extraction from real and synthetic depth images.
• The autoencoder balances reconstruction and classification for mixed-domain data.
• Semantic modeling is applied to the extracted features for an additional level of abstraction.

Main Results:

• The proposed method demonstrates superior performance in cross-modal 3D model retrieval.
• Experimental results on NYUD2 and ModelNet10 datasets validate the effectiveness of the approach.
• The technique successfully retrieves 3D shapes from depth images, outperforming existing methods.

Conclusions:

• The supervised deep autoencoder and semantic modeling approach offers an effective solution for cross-domain 3D model retrieval.
• The method successfully handles noise, occlusion, and view variability in depth images.
• A single deep network trained on mixed data enables effective cross-domain retrieval.