Robustizing Object Detection Networks Using Augmented Feature Pooling

TL;DR: In this paper , a new feature extraction called augmented feature pooling is proposed to improve the robustness of modern object detection algorithms against the large geometric transformation, which integrates the augmented feature maps obtained from the transformed images before feeding it to the detection head without changing the original network architecture.

Abstract: This paper presents a framework to robustize object detection networks against large geometric transformation. Deep neural networks rapidly and dramatically have improved object detection performance. Nevertheless, modern detection algorithms are still sensitive to large geometric transformation. Aiming at improving the robustness of the modern detection algorithms against the large geometric transformation, we propose a new feature extraction called augmented feature pooling. The key is to integrate the augmented feature maps obtained from the transformed images before feeding it to the detection head without changing the original network architecture. In this paper, we focus on rotation as a simple-yet-influential case of geometric transformation, while our framework is applicable to any geometric transformations. It is noteworthy that, with only adding a few lines of code from the original implementation of the modern object detection algorithms and applying simple fine-tuning, we can improve the rotation robustness of these original detection algorithms while inheriting modern network architectures’ strengths. Our framework overwhelmingly outperforms typical geometric data augmentation and its variants used to improve robustness against appearance changes due to rotation. We construct a dataset based on MS COCO to evaluate the robustness of the rotation, called COCO-Rot. Extensive experiments on three datasets, including our COCO-Rot, demonstrate that our method can improve the rotation robustness of state-of-the-art algorithms.