ASK: Adaptively Selecting Key Local Features for RGB-D Scene Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• RGB-D scene classification is challenging due to scattered objects and variable layouts.
• Existing methods struggle with spatial variability and effective local feature extraction.

Purpose of the Study:

• To propose an efficient framework for RGB-D scene recognition that addresses spatial variability.
• To develop a method for adaptively selecting important local features using image labels.

Main Methods:

• Designed a differentiable local feature selection (DLFS) module to extract key scene-related features.
• Utilized spatially-correlated multi-modal RGB-D features and their cross-modal correlations.
• Proposed a variational mutual information maximization loss to ensure discriminative feature selection.

Main Results:

• The DLFS module effectively selects discriminative theme-level and object-level representations.
• The framework leverages RGB-depth correlations for enhanced local feature selection.
• Achieved state-of-the-art performance on public RGB-D scene recognition datasets.

Conclusions:

• The proposed framework efficiently handles spatial variability in RGB-D scene recognition.
• DLFS module enables adaptive selection of local features, improving classification accuracy.
• The method offers a robust solution for complex indoor scene recognition tasks.