Optimizing IoT Video Data: Dimensionality Reduction for Efficient Deep Learning on Edge Computing

This repository contains the official implementation for: "Optimizing IoT Video Data: Dimensionality Reduction for Efficient Deep Learning on Edge Computing". It implements state-of-the-art dimensionality reduction techniques tailored for analyzing bird behaviors from video data, enabling efficient processing on resource-constrained environments like edge computing devices.

Key Highlights

• Dimensionality Reduction Methods:

  • Feature Embeddings: Extract embeddings from pre-trained video models like Swin Transformer, reducing data size by over 6,000 times while maintaining high classification accuracy.
  • Autoencoders: Compress video data using spatio-temporal autoencoders, achieving significant reductions in data size.
  • Single-Frame Analysis: Process individual frames using CNNs, Vision Transformers, or DINO+HoG, leveraging spatial features for classification.

• Dataset:

  • Visual WetlandBirds Dataset: Includes videos of bird behaviors such as feeding, preening, and swimming, annotated with species and actions. The dataset is described in the dataset paper and available on GitHub.

Repository Structure

├── Single_Frame_CNN_Transformer/    # Implementation of single-frame analysis using CNNs and Vision Transformers
├── Single_Frame_Dino_HoG/           # Implementation of single-frame analysis using DINO and HoG features
├── features/                        # Feature embedding extraction from pre-trained video models
├── reduction_autoencoder/           # Implementation of the autoencoder method
├── LICENSE                          
├── README.md