IFeaLiD

IFeaLiD visualizes localization of features in convolutional neural networks by producing similarity-based heat maps to interpret how different layers process image regions for computer vision tasks.

Key Features:

• Interactive Visualization: Enables interactive exploration of high-resolution feature maps with GPU acceleration via WebGL 2 for on-the-fly updates.
• Feature Map Interpretation: Represents neural network layers as multivariate feature maps and visualizes similarity between feature vectors of individual pixels to generate heat maps.
• Layer-Specific Insights: Compares perception of specific image regions to the entire image across network layers to reveal hierarchical feature extraction.
• Real-Time Processing: Processes high-resolution data in real time to support analysis of large-scale computer vision datasets.

Scientific Applications:

• Classification: Interprets convolutional neural network internal representations for image classification tasks.
• Object Detection: Analyzes how network layers encode region-level features relevant to object detection.
• Instance Segmentation: Examines layer-wise feature localization relevant to instance segmentation.

Methodology:

Visualizes feature maps from pre-trained networks such as ResNet101 across computer vision datasets and computes similarity heat maps from pixel-level feature vectors to reveal perceptual differences between regions and layers.