Backtracking Spatial Pyramid Pooling (SPP)-based Image Classifier for Weakly Supervised Top-down Salient Object...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Top-down saliency models typically require extensive pixel-level annotations for training.
• Existing methods often rely on fully supervised learning, limiting their applicability.
• Object detection and localization are crucial tasks in computer vision.

Purpose of the Study:

• To develop a weakly supervised top-down saliency framework for object detection.
• To reduce the reliance on pixel-level annotations by using only binary labels.
• To achieve performance comparable to fully supervised approaches.

Main Methods:

• A backtracking strategy computes the probabilistic contribution of image regions to a CNN classifier's confidence, generating top-down saliency.
• A best-fit bottom-up saliency map is selected and combined with the top-down saliency map.
• Features with high combined saliency train a linear SVM classifier, refined via multi-scale superpixel averaging.

Main Results:

• The proposed weakly supervised framework achieves performance comparable to fully supervised methods.
• Evaluated on seven challenging datasets, demonstrating robust performance.
• Quantitative results show competitive performance against 40 related approaches across four applications.

Conclusions:

• Weakly supervised top-down saliency is a viable and effective alternative to fully supervised methods.
• The framework offers a more efficient approach to saliency map generation for object detection.
• Future work will involve making the code publicly available for broader research use.