Mask-Refined R-CNN: A Network for Refining Object Details in Instance Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Advancements in vision sensors drive progress in computer vision tasks like object detection and tracking.
• Instance segmentation, a complex task, benefits from robust network frameworks such as Mask R-CNN.
• Mask R-CNN's scale-invariant structure overlooks spatial information differences, leading to misclassifications at object edges.

Purpose of the Study:

• To address the limitations of Mask R-CNN in accurately predicting instance details, especially at object boundaries.
• To propose a novel network architecture, Mask-Refined R-CNN (MR R-CNN), for improved instance segmentation.
• To enhance the fusion of global and detailed feature information for more precise segmentation.

Main Methods:

• Modified the stride of Region of Interest Align (ROIAlign) within the network architecture.
• Replaced the standard fully convolutional layer with a semantic segmentation layer.
• Implemented feature fusion using a feature pyramid network and summed forward/backward transmissions of feature maps.

Main Results:

• MR R-CNN achieved approximately 2% higher segmentation accuracy than Mask R-CNN on COCO and Cityscapes datasets.
• Demonstrated a superior average precision of 56.6% for large instances, outperforming existing state-of-the-art methods.
• The proposed method showed low time cost, ease of implementation, and strong generalization ability on the Cityscapes dataset.

Conclusions:

• MR R-CNN effectively overcomes Mask R-CNN's limitations by integrating global and detailed features.
• The enhanced feature fusion strategy significantly improves segmentation accuracy and precision for large instances.
• MR R-CNN presents a computationally efficient and highly generalizable solution for advanced instance segmentation tasks.