Depth Injection Framework for RGBD Salient Object Detection

Area of Science:

• Computer Vision and Image Processing
• Artificial Intelligence
• Machine Learning

Background:

• Accurate salient object detection (SOD) benefits from depth data due to its strong localization capabilities.
• Existing RGBD SOD methods primarily focus on complementary fusion of depth and RGB data.
• A more ambitious integration of depth information into the encoder of a single-stream model is explored.

Purpose of the Study:

• To propose a novel depth injection framework (DIF) for enhanced salient object detection.
• To investigate a new approach of injecting depth maps directly into the encoder of a single-stream model.
• To improve the fusion of RGB and depth data for more accurate SOD.

Main Methods:

• Introduction of a depth injection framework (DIF) comprising an Injection Scheme (IS) and a Depth Injection Module (DIM).
• The IS enhances RGB features by directly injecting depth maps into high-level encoder blocks, maintaining computational efficiency.
• The DIM facilitates cross-modal interaction between depth maps and hierarchical RGB features for effective fusion.

Main Results:

• The proposed DIF achieves state-of-the-art performance on six RGBD datasets for salient object detection.
• The method demonstrates excellent performance on RGB-D SOD tasks.
• The Depth Injection Module (DIM) shows strong generalization, applicable to single-stream SOD models and transformer architectures.

Conclusions:

• The depth injection framework (DIF) offers a powerful and effective method for integrating depth information in SOD.
• The proposed approach significantly enhances salient object detection accuracy and robustness.
• The flexibility of the DIM highlights its potential for broad application in various computer vision models.