SODAR: Segmenting Objects by Dynamically Aggregating Neighboring Mask Representations.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• One-stage instance segmentation models like SOLO offer efficiency but can discard useful neighboring predictions.
• Existing methods often overlook complementary information from adjacent grid cells.

Purpose of the Study:

• To develop a novel learning-based aggregation method (SODAR) to improve SOLO's instance segmentation performance.
• To leverage rich neighboring information while maintaining architectural efficiency.

Main Methods:

• SODAR implicitly learns mask representations encoding geometric structure and context from nearby objects.
• Introduces mask interpolation for efficient sharing of neighboring representations.
• Employs deformable neighbor sampling for adaptive context gathering.

Main Results:

• SODAR significantly improves instance segmentation performance, outperforming SOLO by 2.2 AP on the COCO test set.
• Achieves performance gains with only ~3% additional computation compared to SOLO.
• Demonstrates consistent performance improvements when integrated with the SOLOv2 model.

Conclusions:

• The proposed SODAR method effectively utilizes neighboring information to enhance instance segmentation.
• SODAR offers a computationally efficient approach to boost performance in one-stage instance segmentation models.
• The novel aggregation techniques provide a promising direction for future instance segmentation research.