Mask-Pyramid Network: A Novel Panoptic Segmentation Method.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Existing Mask RCNN methods are computationally intensive due to numerous box proposals and inefficient Non-Maximum Suppression.
• Fusing semantic and instance segmentation results for panoptic segmentation presents challenges.

Purpose of the Study:

• To introduce a novel panoptic segmentation method, the Mask-Pyramid Network.
• To improve computational efficiency and simplify the fusion of semantic and instance segmentation.

Main Methods:

• Proposes a mask pyramid mechanism to generate fewer object proposals by referencing existing segmented masks.
• Generates object proposals and predicts masks hierarchically from larger to smaller sizes.
• Represents object masks as logits for seamless fusion with semantic segmentation logits via SoftMax.

Main Results:

• The Mask-Pyramid Network achieves comparable accuracy to existing methods on Cityscapes and COCO datasets.
• Demonstrates significant computational efficiency compared to traditional approaches.
• Achieves competitive results in panoptic segmentation tasks.

Conclusions:

• The Mask-Pyramid Network offers an efficient and effective solution for panoptic segmentation.
• The proposed method simplifies the fusion of semantic and instance segmentation.
• This approach reduces computational resource consumption while maintaining high accuracy.