Appearance and Pose-Conditioned Human Image Generation Using Deformable GANs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generating realistic person images conditioned on pose and appearance is a challenging task.
• Existing methods struggle with pixel-to-pixel misalignments due to pose variations.
• Accurate person image generation has significant applications, such as in person re-identification.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN) for synthesizing person images with specified pose and appearance.
• To address the challenge of pose-induced pixel misalignments.
• To enhance the performance of person re-identification systems through improved data augmentation.

Main Methods:

• Introduced deformable skip connections within the GAN generator to handle pose differences.
• Proposed a nearest-neighbor loss function for precise detail matching, replacing traditional L1/L2 losses.
• Utilized common datasets and protocols for quantitative and qualitative evaluation.

Main Results:

• The proposed Deformable GANs achieve competitive performance against state-of-the-art methods.
• Generated images preserve visual details of the source appearance while matching the target pose.
• Significantly boosted person re-identification (Re-ID) accuracy when used for data augmentation.

Conclusions:

• Deformable GANs offer a robust solution for conditional person image generation.
• The approach effectively mitigates pose-related alignment issues.
• This method provides superior data augmentation for Re-ID tasks compared to specialized training techniques.