Uncertainty Inspired RGB-D Saliency Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing RGB-D saliency detection models use deterministic approaches, treating the task as point estimation.
• Deterministic solutions are insufficient for capturing the inherent variability in saliency map labeling.

Purpose of the Study:

• To propose the first stochastic framework for RGB-D saliency detection that incorporates uncertainty from the data labeling process.
• To develop a generative architecture for probabilistic RGB-D saliency detection.

Main Methods:

• A generative framework with a generator model (encoder-decoder) and an inference model is proposed.
• The generator maps input images and a latent variable to stochastic saliency predictions.
• Two methods for inferring the latent variable are introduced: Conditional Variational Auto-encoder and Alternating Back-Propagation.

Main Results:

• The proposed framework demonstrates superior performance in learning the distribution of saliency maps.
• Qualitative and quantitative results on six benchmark datasets validate the approach.
• The model effectively utilizes a latent variable to capture labeling variations.

Conclusions:

• The stochastic framework provides a more robust approach to RGB-D saliency detection by accounting for uncertainty.
• The generative architecture and inference methods offer effective solutions for probabilistic saliency prediction.
• The study advances the field by moving beyond deterministic point estimation in saliency detection.