PiCANet: Pixel-wise Contextual Attention Learning for Accurate Saliency Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing saliency models process context holistically, often including irrelevant information.
• Pixel-level saliency prediction can be hindered by noise and distractions within the context region.

Purpose of the Study:

• To propose a novel pixel-wise contextual attention network (PiCANet) for selective context utilization.
• To enhance salient object detection accuracy and uniformity by focusing on informative contextual locations.

Main Methods:

• Developed PiCANet, a network that generates pixel-wise attention maps to select useful contextual features.
• Integrated three PiCANet formulations into a U-Net model for salient object detection, enabling joint training with convolutional neural networks.
• Explored attending to both global and local contexts within the attention mechanism.

Main Results:

• PiCANets effectively incorporate global contrast and regional smoothness for improved salient object localization.
• Demonstrated superior performance in salient object detection compared to state-of-the-art methods.
• Showcased the effectiveness and generalization capabilities of PiCANets on semantic segmentation and object detection tasks.

Conclusions:

• PiCANet offers a significant advancement in saliency detection by enabling pixel-level contextual attention.
• The proposed method enhances the accuracy and robustness of salient object detection and related computer vision tasks.