TinyVision is an innovative research project dedicated to the development of ultra-lightweight image classification models that use minimal parameters. The primary objective is to investigate the essential components required for core vision tasks, achieving this by merging handcrafted feature preprocessing with sophisticated, efficient Convolutional Neural Network (CNN) architectures.

Project Highlights

• Cat vs Dog Classification: The inaugural task utilized a dataset of 25,000 images, employing filter bank preprocessing alongside compact CNNs. Achievements include:
  • Up to 86.87% test accuracy with models limited to 12.5k parameters.
  • Multiple models under 5k parameters demonstrated over 83% accuracy, highlighting impressive efficiency-performance trade-offs.
  • Detailed results and code are available in the cat_vs_dog_classifier/final/v2 directory.
• Cifar10 Classification: The second task employed the Cifar10 dataset, focusing solely on compact CNN architectures without filter bank preprocessing. Results included:
  • A 22.11k parameters model achieving 87.38% accuracy.
  • A 31.15k parameters model attaining 88.43% accuracy.
  • Comprehensive results and code can be found in the cifar10_classifier/final/v1 directory.

Upcoming Features

• Quantization experiments to improve model efficiency.
• In-depth performance analysis of various model architectures to decipher effective design elements.
• Exploration of additional vision tasks such as edge detection and object detection using compact models.
• Expansion of documentation, including architecture diagrams and visualizations for a clearer understanding.
• Reflection on failed or inconclusive experiments, which is vital for delineating design boundaries and enhancing model development.

Project Philosophy

The philosophy underpinning TinyVision emphasizes that small models are not merely designed for speed but also as a fundamental design challenge. The inquiry at the heart of this project is:

How much can we streamline a model without compromising functionality? What components are critical, and what can be considered excess?

By tackling these questions, TinyVision endeavors to create highly efficient models that maintain performance while minimizing complexity.