InstaHMR: Instance-Aware One-Stage Multi-Person Human Mesh Recovery

Area of Science:

• Computer Vision and Machine Learning
• 3D Human Pose and Shape Estimation

Background:

• Estimating 3D human meshes from images is crucial for various applications.
• Existing one-stage methods often struggle with accurate mesh representation, especially in multi-person scenarios.

Purpose of the Study:

• To propose an Instance-aware Multi-person 3D Human Mesh Recovery (InstaHMR) network.
• To improve the accuracy of 3D human mesh recovery by exploiting instance-aware features.

Main Methods:

• Developed a one-stage framework incorporating an instance-aware approach.
• Introduced the Contextual Instance Guidance (CIG) module using spatial and channel attention for instance-specific features.
• Implemented two auxiliary losses: Human Triplet Planes (HTP) loss for joint position accuracy and T-pose Shape (TS) loss for abstract shape learning.

Main Results:

• The proposed CIG module effectively preserves instance-specific information, outperforming pixel-level features.
• Auxiliary losses contribute to capturing subtle joint differences and learning shape parameters.
• Achieved state-of-the-art performance on four benchmark multi-person datasets.

Conclusions:

• InstaHMR demonstrates superior performance in multi-person 3D human mesh recovery.
• The instance-aware feature representation and novel loss functions are key to the model's success.