BRefine: Achieving High-Quality Instance Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Instance segmentation methods like Mask R-CNN achieve high performance but produce coarse masks.
• Downsampling and ROIAlign in backbone networks lead to loss of detailed information, particularly for large objects.
• Low mask resolution results in a sawtooth effect on edges and imprecise segmentation.

Purpose of the Study:

• To propose a novel method, Boundary Refine (BRefine), for achieving high-quality instance segmentation.
• To address the limitations of existing methods in capturing fine details and handling complex boundaries.

Main Methods:

• Utilizes a Fully Convolutional Network (FCN) as the base segmentation architecture.
• Employs a multistage fusion mask head with detailed features to enhance mask resolution.
• Introduces BRank and Sort loss (BR and S loss), combining rank and sort loss with boundary region loss, to tackle segmentation inconsistency and boundary challenges.

Main Results:

• BRefine significantly outperforms Mask R-CNN on COCO, LVIS, and Cityscapes datasets, with improvements of 3.0, 4.2, and 3.5 Average Precision (AP), respectively.
• Demonstrates substantial improvement for segmenting large objects, achieving a 5.0 AP increase on the COCO dataset.
• Successfully handles difficult-to-partition boundaries, producing high-quality segmentation masks.

Conclusions:

• BRefine offers a superior approach to instance segmentation, particularly excelling in mask refinement and boundary accuracy.
• The proposed method effectively mitigates information loss and segmentation inconsistencies inherent in previous architectures.
• BRefine represents a significant advancement in generating precise and detailed instance segmentation masks.