Coreset Selection for Object Detection

核心集选择是一种选择整个数据集中小而代表性子集的方法。它主要在图像分类方面进行研究，假设每个图像只有一个对象。然而，对于物体检测来说，核心集选择更具挑战性，因为一张图像可能包含多个对象。因此，这个主题还需要进行大量的研究。因此，我们介绍了一种新的方法，称为物体检测的核心集选择（CSOD）。CSOD为每个图像中同一类别的多个对象生成图像和类别的代表性特征向量。随后，我们采用子模块化优化来考虑代表性和多样性，并在子模块化优化过程中利用代表性向量来选择子集。当我们在Pascal VOC数据集上评估CSOD时，选择200个图像时，CSOD的AP$_{50}$比随机选择高出6.4个百分点。

CSOD introduces a new approach for coreset selection in object detection, which is more challenging than image classification due to the presence of multiple objects in an image.

CSOD generates imagewise and classwise representative feature vectors for multiple objects of the same class within each image, and adopts submodular optimization for selecting a subset that considers both representativeness and diversity.

CSOD outperformed random selection by +6.4%p in AP$_{50}$ when selecting 200 images on the Pascal VOC dataset. The experiment was designed to evaluate the effectiveness of CSOD and its comparison with random selection. The paper did not mention whether the code is open-source or not.

Previous research has mainly focused on coreset selection in image classification, assuming one object per image. There is limited research on coreset selection for object detection. Some related works include 'Coreset-Based Object Detection' and 'Coreset Methods for Object Detection in Video Surveillance'.