[论文解读] Fully Convolutional Attention Networks for Fine-Grained Recognition

Xiao Liu, Tian Xia+4 more|arXiv (Cornell University)|Mar 22, 2016

Domain Adaptation and Few-Shot Learning参考文献 35被引用 128

一句话总结

FCANs 使用强化学习与一个全卷积网络来定位多個具有辨識性的部件，无需部件注释，使训练/测试更快并在细粒度基准上具有竞争力的准确性。

ABSTRACT

Fine-grained recognition is challenging due to its subtle local inter-class differences versus large intra-class variations such as poses. A key to address this problem is to localize discriminative parts to extract pose-invariant features. However, ground-truth part annotations can be expensive to acquire. Moreover, it is hard to define parts for many fine-grained classes. This work introduces Fully Convolutional Attention Networks (FCANs), a reinforcement learning framework to optimally glimpse local discriminative regions adaptive to different fine-grained domains. Compared to previous methods, our approach enjoys three advantages: 1) the weakly-supervised reinforcement learning procedure requires no expensive part annotations; 2) the fully-convolutional architecture speeds up both training and testing; 3) the greedy reward strategy accelerates the convergence of the learning. We demonstrate the effectiveness of our method with extensive experiments on four challenging fine-grained benchmark datasets, including CUB-200-2011, Stanford Dogs, Stanford Cars and Food-101.

研究动机与目标

• Motivate fine-grained recognition where small inter-class differences and large intra-class variations exist.
• Eliminate dependence on expensive ground-truth part annotations by using weakly supervised learning.
• Propose a fully convolutional attention framework that reuses feature maps for efficiency during training and testing.
• Enable localization of multiple discriminative parts with greedy, step-wise rewards to accelerate training.

提出的方法

• Propose FCANs consisting of a shared feature network, an attention network producing multiple part score maps, and a per-part classification network.
• Use a Markov Decision Process formulation where actions are attention locations and rewards reflect classification quality.
• Train with REINFORCE-based policy gradients using a greedy reward strategy that grants intermediate rewards when accuracy improves.
• Reuse convolutional feature maps across time steps to avoid recomputing features (Fast-RCNN-like sharing).
• Crop high-resolution regions around attended locations for final classification while keeping a shared representation for efficiency.

实验结果

研究问题

• RQ1 Can weakly supervised attention learn discriminative parts for fine-grained recognition without part annotations?
• RQ2 Does a fully convolutional attention architecture improve efficiency over recurrent attention models while maintaining accuracy?
• RQ3 How many attentions and what reward strategy yield the best accuracy and training convergence across datasets?

主要发现

Dataset	Accuracy (%)
CUB-200-2011	84.3
Stanford Dogs	88.9
Stanford Cars	91.5
Food-101	86.3

• Achieves competitive fine-grained accuracy on four benchmarks without using part annotations at test time.
• Outperforms prior RL-based attention models in both accuracy and efficiency due to fully convolutional feature reuse.
• Two attentions provide a good trade-off between accuracy and computational cost, with diminishing gains beyond two attentions.
• Greedy reward strategy accelerates training convergence and improves final accuracy compared to only-end rewards.
• Training with shared feature maps and Fast-RCNN-like region extraction significantly reduces computation and speeds up testing.