KepSalinst: Using Peripheral Points to Delineate Salient Instances

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Salient instance segmentation (SIS) is an advancement from salient object detection (SOD), focusing on segmenting individual salient objects.
• Dynamic convolutions have shown promise in various segmentation tasks.
• Existing methods may lack precise spatial delineation and saliency awareness.

Purpose of the Study:

• To introduce a novel keypoints-based salient instance segmentation network (KepSalinst).
• To leverage geometrical guidance from instance keypoints for dynamic convolutions.
• To enhance saliency perception and instance-level scoring.

Main Methods:

• Developed KepSalinst, a network utilizing center and peripheral keypoints for dynamic convolution guidance.
• Designed a differentiated patterns fusion (DPF) module to integrate complementary features from keypoints.
• Introduced a high-level semantic guided saliency (HSGS) module for instance saliency scoring.

Main Results:

• KepSalinst demonstrated superior performance across four benchmark SIS datasets (ILSO, SOC, SIS10K, COME15K).
• The method achieved both qualitative and quantitative improvements over existing approaches.
• Keypoint guidance and fusion modules effectively improved segmentation accuracy.

Conclusions:

• KepSalinst offers a robust framework for salient instance segmentation.
• The integration of keypoint geometry and semantic saliency significantly boosts performance.
• This work advances the state-of-the-art in salient instance segmentation.