Dynamic Filtering with Large Sampling Field for ConvNets.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (ConvNets) are fundamental in computer vision.
• Existing methods like Dynamic Filtering Networks (DFN) avoid feature map blurring but can be prone to overfitting.
• There is a need for efficient ConvNet strategies that expand receptive fields without excessive parameter growth.

Purpose of the Study:

• To propose a novel dynamic filtering strategy with a large sampling field for ConvNets (LS-DFN).
• To enhance feature learning by incorporating information from sampled neighboring regions.
• To address the overfitting issue in models with increased parameters while maintaining efficiency.

Main Methods:

• Developed LS-DFN, a strategy where position-specific kernels learn from multiple sampled neighbor regions.
• Integrated residual learning to facilitate training.
• Applied an attention mechanism for effective fusion of features from different samples.
• Ensured end-to-end training via standard back-propagation.

Main Results:

• LS-DFN significantly enlarges kernel receptive fields without a proportional increase in parameters.
• The model inherits DFN's advantages: avoiding feature map blurring and maintaining translation invariance.
• Demonstrated superior performance on object detection and semantic segmentation tasks (VOC benchmark).
• Achieved sharper responses in flow estimation (FlyingChairs dataset).

Conclusions:

• LS-DFN effectively balances expanded receptive fields with parameter efficiency, mitigating overfitting.
• The proposed method shows strong recognition and response capabilities across various computer vision tasks.
• LS-DFN offers a promising approach for improving ConvNet performance in sparse and dense prediction tasks.