ROM-Pose: restoring occluded mask image for 2D human pose estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Human Pose Estimation (HPE) is crucial for understanding human actions in images.
• Top-down HPE methods rely on accurate bounding box detection, which is challenged by occlusions.
• Occluded body parts are often excluded from HPE model input, reducing accuracy.

Purpose of the Study:

• To address inaccuracies in top-down human pose estimation caused by bounding box detection failures due to occlusions.
• To introduce a novel method, ROM-Pose, that restores occluded image regions to improve HPE.
• To enhance the input for HPE models by including information from previously occluded areas.

Main Methods:

• Developed ROM-Pose, a system combining a restoration model and an HPE model.
• The restoration model uses a specialized Whole Common Objects in Context (COCO) dataset to identify and restore occluded image sections.
• Restored images are integrated with original RGB images to provide complete input for the HPE model.

Main Results:

• ROM-Pose successfully restores occluded image parts, enabling more comprehensive bounding box detection.
• The enhanced input allows the HPE model to recognize previously excluded body parts.
• Achieved a 1.6% improvement in average precision (AP) compared to the baseline HPE method.

Conclusions:

• ROM-Pose effectively mitigates the negative impact of occlusions on human pose estimation accuracy.
• Restoring occluded image information is a viable strategy to improve bounding box detection in HPE.
• The proposed method offers a significant advancement for robust human pose estimation in complex visual scenes.