End-to-end human parsing and detection optimized for resource-constrained devices.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human parsing is crucial for human-centric analysis, segmenting clothing and body parts.
• Current methods often require auxiliary inputs, hindering use on resource-constrained devices.

Purpose of the Study:

• To propose an end-to-end framework for human parsing optimized for low-resource environments.
• To enhance contextual understanding and enable simultaneous detection and parsing.

Main Methods:

• Implemented a transformer-based self-attention module for improved contextual information.
• Introduced bounding-polygon annotations for integrated detection and parsing.
• Optimized the framework for efficient inference on resource-limited hardware.

Main Results:

• Achieved fine-grained human parsing results in a single inference pass.
• Significantly improved inference speed without compromising accuracy.
• Demonstrated effectiveness and efficiency on a Raspberry Pi for real-world validation.

Conclusions:

• The proposed framework offers an efficient and accurate solution for human parsing in resource-constrained settings.
• The integration of self-attention and bounding-polygon annotations advances single-pass human parsing capabilities.