Multi-Horizon Feature Reconstruction for 3D Salient Object Detection

The recent improvements of current RGB-D salient object detection models have achieved better results by leveraging the depth modality in a convolutional neural network. Most existing approaches use a multi-stream architecture with enhanced feature fusion to subsequently determine saliency. The inconsistency of salient regions may result from the limitations of the acquisition devices and the low-depth images. Multi-horizon feature characteristics and complementary modalities are essential for resolving these problems. Even in cluttered backgrounds, complementary and multi-horizon features across several horizons must be taken into consideration since they provide structural and contextual information about the salient object. To maximize important features in low-depth images, the proposed model addresses these constraints. The multi-horizon features are optimized from the RGB and depth streams by using the proposed backbone network with the feature aggregation and optimization (FAO) module. The FAO module, which optimizes feature extraction by utilizing complementary features from the depth stream, is used to enhance the RGB stream. The proposed model investigates multi-horizon features between multi-resolution and multi-stage complementary features using a multi-horizon reconstruction module. To forecast optimal saliency, the FAO module combines data at multiple levels. A guiding framework for maximizing salient regions and reducing non-salient ones throughout the fusion process is provided by attention maps at various phases. Current evaluation metrics from the six publicly accessible difficult RGB-D datasets are used in the experimental investigation. Furthermore, we compare our findings with those of 18 other cutting-edge methods that have demonstrated encouraging performance on the above challenges.