SAVE: Self-Attention on Visual Embedding for Zero-Shot Generic Object Counting

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generic Visual Object Counting aims to identify and quantify objects within images.
• Zero-shot counting enables object counting for arbitrary classes without prior examples, contrasting with few-shot methods that require exemplars.
• Existing methods often require exemplars or lack the automation needed for rapid processing.

Purpose of the Study:

• To propose a fully automated zero-shot counting method that outperforms current zero-shot and few-shot approaches.
• To enhance the accuracy and efficiency of visual object counting across various domains.

Main Methods:

• Exploiting feature maps from a pre-trained detection-based backbone.
• Introducing a Visual Embedding Module to generate semantic embeddings with object contextual information.
• Utilizing a Self-Attention Matching Module to create an encoded representation for the head counter.

Main Results:

• Achieved state-of-the-art performance in zero-shot counting on the FSC147 dataset.
• Obtained the best Mean Absolute Error (MAE) of 8.89 and Root Mean Square Error (RMSE) of 35.83.
• Demonstrated competitive results compared to few-shot methods.

Conclusions:

• The proposed method offers a significant advancement in automated zero-shot visual object counting.
• The approach shows promise for applications in tree counting, wildlife monitoring, and medical image analysis (e.g., blood cell counting).
• This work pushes the boundaries of visual object counting, enabling more efficient and accurate automated solutions.