Leveraging Instance-, Image- and Dataset-Level Information for Weakly Supervised Instance Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for semantic instance segmentation typically require pixel-wise masks or bounding box annotations, which are costly to obtain.
• Weakly supervised learning aims to reduce annotation effort by using less precise labels, such as image-level tags.

Purpose of the Study:

• To develop a method for weakly supervised semantic instance segmentation using only image-level supervision.
• To alleviate the data-hungry nature of deep learning models in computer vision tasks.

Main Methods:

• Aggregating image-level information from all training images into a knowledge graph to exploit semantic relationships.
• Employing a multiple instance learning (MIL) framework trained end-to-end with image-level labels.
• Utilizing segment-based object proposals (SOPs) and formulating a large undirected graph for optimal multi-way cut to assign category labels.

Main Results:

• The proposed method generates denoised SOPs with assigned category labels, serving as pseudo instance segmentations.
• These pseudo instance segmentations are used to train fully supervised models.
• The approach achieves state-of-the-art performance in both weakly supervised instance segmentation and semantic segmentation tasks.

Conclusions:

• The developed knowledge graph aggregation and MIL framework effectively address the challenge of semantic instance segmentation with limited supervision.
• This method significantly reduces the reliance on expensive annotations, making deep learning models more data-efficient.
• The code is publicly available, facilitating further research and application.