Light Field Saliency Detection with Deep Convolutional Networks.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Light field imaging captures light direction, offering advantages over RGB imaging for tasks like saliency detection.
• Convolutional Neural Networks (CNNs) excel at saliency detection in RGB images but require adaptation for light field data.
• Existing light field datasets are insufficient for training deep CNNs.

Purpose of the Study:

• To address limitations in light field saliency detection by introducing a new dataset and a CNN framework.
• To facilitate the training of deep networks and benchmarking of light field saliency detection methods.
• To develop a CNN-based approach capable of processing the unique angular patterns in light field images.

Main Methods:

• Creation of the Lytro Illum dataset, comprising 640 light fields and saliency maps.
• Development of a novel end-to-end CNN framework incorporating three MAC (Model Angular Changes) blocks.
• Systematic evaluation of network architecture variants and comparison with 2D saliency detection.

Main Results:

• The Lytro Illum dataset is larger, higher quality, and more varied than existing datasets.
• The proposed CNN framework significantly outperforms state-of-the-art methods on the new dataset.
• The framework demonstrates strong generalization capabilities on other light field datasets.

Conclusions:

• The developed dataset and CNN framework advance the field of light field saliency detection.
• The novel MAC blocks effectively process angular information in light fields.
• The proposed method offers a robust solution for salient region detection in light field imagery.