Source-Free Model Transferability Assessment for Smart Surveillance via Randomly Initialized Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Smart City Technologies

Background:

• Smart surveillance cameras are vital for automated tasks in smart cities.
• A universal model is suboptimal due to diverse deployment environments.
• Selecting the right model for new sites is challenging without labeled data.

Purpose of the Study:

• To develop an automated framework for assessing model transferability in smart surveillance.
• To identify the most suitable pre-trained model for a new deployment site.
• To overcome the challenge of unavailable source and labeled target data.

Main Methods:

• Constructing a model zoo of diverse environmental context models.
• Leveraging embeddings from randomly initialized neural networks (RINNs) for task-agnostic reference embeddings.
• Quantifying embedding similarity using minibatch-Centered Kernel Alignment (CKA) without pretraining.

Main Results:

• High correlations between the embedding-level score and ground-truth model rankings (Kendall's τ: 0.95, 0.94, 0.89).
• Demonstrated consistent selection of the most transferable model across various downstream tasks.
• Validated the framework's practicality for object tagging, anomaly detection, and event classification.

Conclusions:

• The proposed framework offers a robust, low-cost solution for selecting optimal surveillance models.
• It effectively addresses the need for model adaptation or retraining in smart city deployments.
• The method eliminates the need for labeled data, enhancing practical applicability.