Few-Shot Object Recognition from Machine-Labeled Web Images

TL;DR

This paper introduces an abstraction memory framework for few-shot object recognition that leverages large-scale machine-labeled web images, achieving near-human performance with minimal labeled data.

Contribution

It presents a novel abstraction memory model utilizing machine-labeled datasets for effective few-shot object recognition, reducing reliance on human annotations.

Findings

Model performs well on ImageNet with human labels.

Machine-labeled annotations nearly match human-labeled data in accuracy.

Effective use of external memory improves few-shot learning performance.

Abstract

With the tremendous advances of Convolutional Neural Networks (ConvNets) on object recognition, we can now obtain reliable enough machine-labeled annotations easily by predictions from off-the-shelf ConvNets. In this work, we present an abstraction memory based framework for few-shot learning, building upon machine-labeled image annotations. Our method takes some large-scale machine-annotated datasets (e.g., OpenImages) as an external memory bank. In the external memory bank, the information is stored in the memory slots with the form of key-value, where image feature is regarded as key and label embedding serves as value. When queried by the few-shot examples, our model selects visually similar data from the external memory bank, and writes the useful information obtained from related external data into another memory bank, i.e., abstraction memory. Long Short-Term Memory (LSTM) controllers and attention mechanisms are utilized to guarantee the data written to the abstraction memory is correlated to the query example. The abstraction memory concentrates information from the external memory bank, so that it makes the few-shot recognition effective. In the experiments, we firstly confirm that our model can learn to conduct few-shot object recognition on clean human-labeled data from ImageNet dataset. Then, we demonstrate that with our model, machine-labeled image annotations are very effective and abundant resources to perform object recognition on novel categories. Experimental results show that our proposed model with machine-labeled annotations achieves great performance, only with a gap of 1% between of the one with human-labeled annotations.