Partial Convolution for Padding, Inpainting, and Image Synthesis.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Standard convolutions can introduce artifacts in corrupted images or during conditional synthesis.
• Existing padding schemes in convolutional networks make strong assumptions, potentially impairing model accuracy.
• Partial convolution selectively processes valid pixels using binary masks, addressing limitations of standard convolutions.

Purpose of the Study:

• To explore and unify the applications of partial convolution.
• To investigate the efficacy of partial convolution-based padding in various computer vision tasks.
• To demonstrate the advantages of partial convolution over traditional padding methods.

Main Methods:

• Reviewing existing partial convolution applications.
• Implementing and evaluating partial convolution-based padding.
• Conducting comprehensive studies on image classification, 3D action recognition, and semantic segmentation.

Main Results:

• Partial convolution-based padding demonstrates consistent improvements across diverse computer vision tasks.
• The proposed padding method outperforms strong baselines in evaluated tasks.
• Partial convolution effectively handles valid and corrupted pixels, reducing artifacts.

Conclusions:

• Partial convolution-based padding is a promising technique for enhancing convolutional neural network performance.
• This approach offers a more robust alternative to conventional padding schemes.
• Further exploration of partial convolution in computer vision is warranted.