Not All Samples are Trustworthy: Towards Deep Robust SVP Prediction

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Estimating subjective visual properties (SVP) from images is an emerging computer vision task.
• Crowdsourced datasets for SVP often contain noise due to a lack of quality control, leading to untrustworthy samples.

Purpose of the Study:

• To develop robust models capable of learning from noisy crowdsourced annotations for SVP estimation.
• To construct general learning frameworks that address the challenge of unreliable data in SVP datasets.

Main Methods:

• Proposed a probabilistic framework to explicitly model sparse unreliable patterns in SVP datasets.
• Developed an alternative framework reformulating unreliable patterns as a "contraction" operation on the loss function, enabling efficient end-to-end training and theoretical analysis.
• Implemented frameworks with models interpreting sparse noise parameters using HodgeRank theory.

Main Results:

• Demonstrated the effectiveness of the proposed robust learning frameworks through extensive theoretical and empirical studies.
• Showcased improved performance in estimating subjective visual properties from noisy crowdsourced data.

Conclusions:

• The developed frameworks provide effective solutions for learning from noisy crowdsourced annotations in SVP estimation.
• The proposed methods enhance the reliability and accuracy of subjective visual property estimation in computer vision applications.