SCG: Saliency and Contour Guided Salient Instance Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient instance segmentation (SIS) presents unique challenges, including segmenting only salient objects while ignoring backgrounds and handling generic object instances without predefined categories.
• Existing methods often struggle with these dual challenges, necessitating advancements beyond conventional instance segmentation techniques.

Purpose of the Study:

• To address the difficulties in salient instance segmentation (SIS) by proposing a novel approach that leverages complementary saliency and contour information.
• To enhance the state-of-the-art Mask R-CNN model for improved performance in segmenting salient, generic object instances.

Main Methods:

• An improved Mask R-CNN architecture featuring an interleaved execution strategy and a novel mask head network for incorporating global context.
• Addition of dedicated saliency and contour detection branches to Mask R-CNN, fusing their features with the base model's features.
• Development of a multiscale global attention model for generating attentive global features from multiscale representative features to facilitate feature fusion.

Main Results:

• All proposed model components demonstrated improvements in salient instance segmentation (SIS) performance.
• The overall model outperformed existing state-of-the-art SIS methods by over 6% and Mask R-CNN by over 3%.
• Utilizing additional multitask training data further enhanced model performance on the ILSO dataset.

Conclusions:

• The proposed method effectively addresses the challenges of salient instance segmentation by integrating saliency and contour cues.
• The novel architecture and attention mechanism significantly advance the capabilities of generic object instance segmentation.
• The findings indicate a promising direction for future research in robust and versatile instance segmentation models.