Few-Shot Deep Adversarial Learning for Video-based Person Re-identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video-based person re-identification (re-ID) faces challenges with scalability due to extensive labeling requirements for existing supervised methods.
• Learning view-invariant video representations and capturing temporal dynamics are crucial for effective person re-ID, especially with limited training samples.

Purpose of the Study:

• To propose a novel few-shot deep learning approach for video-based person re-ID.
• To develop a method that learns comparable, discriminative, and view-invariant representations from video data.
• To address the limitations of existing methods in terms of scalability and the explicit learning of view-invariant features.

Main Methods:

• A few-shot deep learning approach utilizing variational recurrent neural networks (VRNNs).
• Adversarial training to generate latent variables with temporal dependencies.
• Focus on learning representations that are both discriminative and view-invariant.

Main Results:

• The proposed method successfully creates view-invariant temporal features from video footages.
• Empirical evidence from three benchmark datasets demonstrates state-of-the-art performance.
• The approach shows effectiveness in matching persons across different camera views with limited supervision.

Conclusions:

• The developed method offers a scalable and effective solution for video-based person re-ID.
• The VRNN-based adversarial approach enables learning of robust, view-invariant representations.
• This work advances the field by achieving high performance with few-shot learning in person re-ID.