Pixel is All You Need: Adversarial Spatio-Temporal Ensemble Active Learning for Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly-supervised learning reduces data annotation costs but its efficacy for saliency models compared to fully-supervised methods remains unclear.
• Existing active learning methods can be overconfident, hindering accurate pixel identification for training.
• Saliency model training requires extensive, densely annotated datasets, which are labor-intensive to create.

Purpose of the Study:

• To investigate if saliency models trained with point-annotated data can achieve performance equivalent to fully-supervised training.
• To propose and validate a novel active learning framework for efficient saliency model training.
• To provide theoretical proof for the existence of effective point-labeled datasets for saliency modeling.

Main Methods:

• Developed an adversarial attack triggering uncertainty to identify uncertain pixels and overcome model overconfidence.
• Implemented a spatio-temporal ensemble strategy to enhance performance and reduce computational cost.
• Introduced relationship-aware diversity sampling to prevent oversampling and improve model accuracy.

Main Results:

• The proposed method successfully identified a point-labeled dataset enabling equivalent performance to fully-supervised models.
• Saliency models trained on the derived point-labeled dataset achieved 98%-99% of the performance of fully-supervised versions.
• The approach required only ten annotated points per image, demonstrating significant efficiency gains.

Conclusions:

• Weakly-supervised saliency modeling with point annotations can achieve performance comparable to fully-supervised methods.
• The novel adversarial spatio-temporal ensemble active learning framework effectively reduces annotation requirements.
• This research validates the hypothesis that high-performance saliency models can be trained with minimal labeled data.