GenDet: Meta Learning to Generate Detectors From Few Shots
IEEE Transactions on Neural Networks and Learning Systems
|February 1, 2021
Summary
GenDet, a novel meta-learning framework, explicitly generates object detectors for novel classes using few shots. This approach significantly improves few-shot object detection performance on benchmark datasets.
Area of Science:
- Computer Vision
- Machine Learning
Background:
- Object detection models struggle with novel classes lacking sufficient training data.
- Existing few-shot detection methods often address the problem implicitly.
Purpose of the Study:
- To propose GenDet, a meta-learning framework for explicit few-shot object detection.
- To enable effective generation of object detectors for novel classes with limited data.
Main Methods:
- GenDet trains a detector generator using numerous few-shot tasks from base classes.
- Incorporates an adaptive pooling module to refine generated detectors.
- Utilizes reference detectors for base classes to guide generator training.
- Encourages orthogonality among generated detectors for enhanced generalization.
Main Results:
- GenDet achieves state-of-the-art results on ImageNet, VOC, and COCO datasets.
- Demonstrates superior performance in various few-shot object detection scenarios.
- Effectively generates detectors for novel classes from limited examples.
Conclusions:
- GenDet offers an effective explicit approach to few-shot object detection.
- The meta-learning framework shows strong generalization capabilities for novel classes.
- The method sets a new benchmark for few-shot object detection tasks.
