Self-Correction for Human Parsing | JoVE Visualize

Area of Science:

Computer Vision
Machine Learning
Artificial Intelligence

Background:

Pixel-level mask labeling for fine-grained semantic segmentation, particularly human parsing, is hindered by ambiguous boundaries and similar appearances, leading to noisy ground-truth labels.
Label noise significantly degrades the training process and performance of semantic segmentation models.
Existing methods often struggle to effectively address label inaccuracies in complex datasets.

Purpose of the Study:

To develop a noise-tolerant method for human parsing that enhances the reliability of supervised labels and improves model performance.
To introduce Self-Correction for Human Parsing (SCHP), a technique designed to progressively refine both labels and models through iterative learning.
To demonstrate the model-agnostic nature of SCHP and its applicability to various human parsing architectures.

Main Methods:

Introduced Self-Correction for Human Parsing (SCHP), a noise-tolerant approach for semantic segmentation.
Designed a cyclically learning scheduler to infer reliable pseudo masks by iteratively aggregating current and former sub-optimal models.
Employed a self-correction mechanism where refined labels further boost model performance in a feedback loop.

Main Results:

Achieved new state-of-the-art results on 6 benchmarks across single, multi-person, and video human parsing tasks (LIP, Pascal-Person-Part, ATR, CIHP, MHP, VIP).
Demonstrated the effectiveness of SCHP across four different human parsing models (Deeplab V3+, CE2P, OCR, CE2P+).
Secured 1st place in all three human parsing tracks at the 3rd Look Into Person Challenge.

Conclusions:

SCHP effectively mitigates the negative impact of label noise in human parsing datasets.
The proposed method enables reciprocal improvement between model accuracy and label reliability through self-correction cycles.
SCHP offers a robust and versatile solution for enhancing semantic segmentation performance, validated by extensive experiments and competitive results.