Weakly Supervised Salient Object Detection by Learning A Classifier-Driven Map Generator.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Top-down saliency detection typically requires extensive pixel-wise annotated data.
• The high cost of data collection limits the application of traditional saliency detection methods.
• Weakly supervised learning offers a potential solution by utilizing less detailed annotations.

Purpose of the Study:

• To develop a weakly supervised framework for object saliency detection.
• To reduce the reliance on pixel-wise annotated training data.
• To achieve competitive or superior performance compared to fully supervised methods.

Main Methods:

• A framework with two collaborative Convolutional Neural Network (CNN) modules: an image-level classifier and a pixel-level map generator.
• Utilizing image-level labels (presence/absence of target object) for training.
• Incorporating background prior, superpixel, and object proposal evidence into the map generator.

Main Results:

• The proposed method successfully generates saliency maps using only image-level labels.
• The integration of multiple cues (background, superpixels, object proposals) regularizes training and prevents overfitting.
• Experimental results demonstrate superior performance, even surpassing fully supervised approaches.

Conclusions:

• Weakly supervised object saliency detection is feasible and effective.
• The proposed framework offers a cost-efficient alternative to traditional methods.
• The method shows strong potential for real-world applications where pixel-level annotations are scarce.