Fully Convolutional Networks for Panoptic Segmentation With Point-Based Supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Panoptic segmentation combines instance segmentation and semantic segmentation.
• Existing methods often require complex architectures or extensive annotations.

Purpose of the Study:

• To develop a unified, efficient framework for fully- and weakly-supervised panoptic segmentation.
• To introduce a novel, cost-effective point-based annotation method for weak supervision.

Main Methods:

• Panoptic FCN utilizes a fully convolutional pipeline with a kernel generator.
• Predictions are made by convolving high-resolution features directly.
• A new point-based annotation strategy is proposed for weak supervision.

Main Results:

• Panoptic FCN outperforms previous box-based and box-free models in efficiency and performance.
• Achieved 82% of fully-supervised performance with only 20 random points per instance in the weakly-supervised setting.
• Established new benchmarks on COCO, VOC 2012, Cityscapes, and Mapillary Vistas datasets.

Conclusions:

• Panoptic FCN offers a simple, strong, and efficient solution for panoptic segmentation.
• The proposed weak supervision method significantly reduces annotation costs.
• Panoptic FCN demonstrates superior performance and efficiency in both fully- and weakly-supervised scenarios.