In defense of local descriptor-based few-shot object detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• State-of-the-art image object detection models require extensive training data and parameter tuning.
• Human intelligence excels at few-shot learning, recognizing new concepts from minimal examples.
• Classical hand-crafted local descriptors lack global structural understanding, limiting their performance.

Purpose of the Study:

• To develop a novel few-shot object detection approach that overcomes limitations of traditional methods.
• To enhance local descriptors by incorporating global context and semantic information.
• To achieve robust object detection with minimal training data, inspired by human perception.

Main Methods:

• Refined classical local descriptors (e.g., SIFT, HOG) with spatial contextual attention and neighbor affinities.
• Embedded local descriptors into a discriminative subspace using Kernel-InfoNCE loss.
• Developed a brain-inspired, few-shot feature representation combining primitive representation and semantic context learning.

Main Results:

• The proposed method achieves effective few-shot object detection with a nearly learning-free approach.
• Experiments on remote sensing imageries demonstrate the model's efficacy in 2-D affine spaces.
• The approach enables accelerated, non-parametric visual similarity computation for object detection.

Conclusions:

• Integrating global structure sense into local descriptors significantly improves few-shot object detection performance.
• The brain-inspired feature representation facilitates generalization and robust learning from few examples.
• This nearly learning-free method offers an efficient alternative for object detection in data-scarce scenarios, particularly in remote sensing.