Data-Level Recombination and Lightweight Fusion Scheme for RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing RGB-D salient object detection methods use a bistream architecture, processing RGB and depth data separately.
• This approach leads to performance bottlenecks as depth data is often less informative than RGB data, biasing fusion towards RGB.
• The conventional bistream architecture struggles to effectively fuse complementary information from both RGB and depth modalities.

Purpose of the Study:

• To propose a novel data-level recombination strategy for fusing RGB and depth (D) data before deep feature extraction.
• To develop a lightweight triple-stream network to optimally fuse channel-wise complementary information.
• To overcome the performance limitations of existing bistream architectures in salient object detection.

Main Methods:

• Cyclically converted the 4-dimensional RGB-D data into DGB, RDB, and RGD formats.
• Applied a newly designed lightweight triple-stream network to the novel data formulations.
• Achieved optimal channel-wise complementary fusion between RGB and depth features.

Main Results:

• The proposed data-level recombination strategy and triple-stream network achieved a new state-of-the-art (SOTA) performance.
• Demonstrated superior salient object detection compared to existing bistream methods.
• Showcased the effectiveness of fusing modalities at the data level before feature extraction.

Conclusions:

• The novel data recombination strategy effectively addresses the information imbalance between RGB and depth data.
• The lightweight triple-stream network provides an optimal fusion mechanism for enhanced salient object detection.
• This approach sets a new benchmark for RGB-D salient object detection performance.