Location Sensitive Network for Human Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Instance segmentation is crucial for distinguishing individual objects in images.
• Location information is a key feature for accurate instance segmentation.
• Existing methods often struggle with distinguishing heavily occluded instances.

Purpose of the Study:

• To propose a novel model, the Location Sensitive Network (LSNet), for enhanced human instance segmentation.
• To integrate instance-specific location information effectively into a one-stage segmentation framework.
• To improve the model's ability to handle severe occlusion cases.

Main Methods:

• LSNet utilizes a Pose Attention Module (PAM) for coordinate encoding to integrate location information into attention regions.
• The Keypoints Sensitive Combination (KSCom) operation leverages multiple sampling points, including human keypoints and random points, for instance representation.
• PAM and KSCom work together to enable feature-level instance distinction and reduce misclassified pixels.

Main Results:

• LSNet-5 achieved a mean Average Precision (mAP) of 56.2 at 18.5 Frames Per Second (FPS) on the COCOPersons dataset.
• The proposed method demonstrates significant superiority over existing approaches, particularly in scenarios with severe occlusion.
• Experiments on public datasets validate the effectiveness of the LSNet model.

Conclusions:

• LSNet provides an effective approach for human instance segmentation by incorporating location-aware mechanisms.
• The integration of location information via PAM and KSCom significantly enhances instance discrimination, especially under occlusion.
• The model's performance indicates a promising advancement in the field of instance segmentation.