Revisiting Feature Acquisition Bias for Few-Shot Fine-Grained Image Classification

Weichuan Zhang, Jiapan Guo, Ziqiang Zheng, Yongsheng Gao, Changming Sun

Abstract: Recent work on metric-learning based few-shot fine-grained image classification (FSFGIC) has achieved promising success in classification accuracy, where various convolution neural networks with different similarity measures are utilized to learn a common feature representation for each category for FSFGIC. In this paper, we identify and analyze for the first time a fundamental problem of existing metric-learning based FSFGIC methods which fail to effectively address the bias in features obtained from each input image that causes misclassification. To solve this problem, we present a robust feature acquisition network (RFANet) that has the ability to effectively address the bias in the feature information obtained from each input image and guide convolution-based embedding models to significantly increase the accuracy. Our proposed architecture can be easily embedded into any episodic training mechanisms for end-to-end training from scratch. Extensive experiments on FSFGIC tasks demonstrate the superiority of the proposed method over the state-of-the-arts.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning