Probabilistic Graph Attention Network With Conditional Kernels for Pixel-Wise Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel at pixel-level prediction using multi-scale representations.
• Existing methods often fuse features through simple averaging or concatenation, limiting performance.
• A principled approach for structured multi-scale feature learning and fusion is needed.

Purpose of the Study:

• To develop a novel approach for structured multi-scale feature learning and fusion.
• To advance the state-of-the-art in pixel-level prediction tasks.
• To improve the learning capacity and fusion of multi-scale representations.

Main Methods:

• Proposed a probabilistic graph attention network with Attention-Gated Conditional Random Fields (AG-CRFs).
• Introduced feature-dependent conditional kernels within a deep probabilistic framework.
• Evaluated on BSDS500, NYUD-V2, KITTI, and Pascal-Context datasets.

Main Results:

• Demonstrated the effectiveness of the proposed latent AG-CRF model.
• Achieved state-of-the-art results in challenging pixel-wise prediction problems.
• Validated the approach on monocular depth estimation, object contour prediction, and semantic segmentation.

Conclusions:

• The proposed probabilistic graph attention network with feature conditional kernels effectively learns and fuses structured multi-scale features.
• The AG-CRF model offers a principled way to enhance pixel-level prediction.
• The method shows significant improvements across diverse datasets and tasks.