Retail50K

鼓励社区为更复杂的环境调整定向边界框 (OBB) 检测器的数据集。使用定向边界框 (OBB) 的对象检测可以通过减少与背景区域的重叠来更好地定位旋转对象。现有的 OBB 方法大多建立在水平边界框检测器的基础上，通过引入由距离损失优化的额外角度维度。然而，由于距离损失仅使 OBB 的角度误差最小化，并且它与 IoU 的相关性松散，因此它对高纵横比的物体不敏感。因此，一种新颖的损失，Pixels-IoU (PIoU) Loss，被制定为利用角度和 IoU 来实现准确的 OBB 回归。 PIoU 损失源自 IoU 度量，采用逐像素形式，简单且适用于水平和定向边界框。为了证明其有效性，我们评估了基于锚和无锚框架的 PIoU 损失。实验结果表明，PIoU 损失可以显着提高 OBB 检测器的性能，尤其是在具有高纵横比和复杂背景的物体上。此外，之前的评估数据集不包括对象具有高纵横比的场景，因此引入了新的数据集 Retail50K，以鼓励社区将 OBB 检测器用于更复杂的环境。

This dataset encourages the community to adapt oriented bounding box (OBB) detectors for more complex environments. Object detection using oriented bounding boxes (OBBs) enables better localization of rotated objects by reducing overlap with background regions. Most existing OBB methods are built upon horizontal bounding box detectors by introducing an additional angular dimension optimized via distance loss. However, since distance loss only minimizes the angular error of OBBs and has loose correlation with IoU, it is insensitive to objects with high aspect ratios. Therefore, a novel loss, Pixels-IoU (PIoU) Loss, is formulated to leverage both angle and IoU for accurate OBB regression. PIoU Loss is derived from the IoU metric, adopts a per-pixel form, and is simple and applicable to both horizontal and oriented bounding boxes. To validate its effectiveness, we evaluated PIoU Loss on both anchor-based and anchor-free frameworks. Experimental results demonstrate that PIoU Loss can significantly improve the performance of OBB detectors, especially for objects with high aspect ratios and complex backgrounds. Furthermore, previous evaluation datasets do not cover scenarios where objects have high aspect ratios, so a new dataset, Retail50K, is introduced to encourage the community to apply OBB detectors to more complex environments.

搜集汇总

背景与挑战

背景概述

Retail50K是一个专注于定向边界框（OBB）检测的数据集，特别针对高纵横比和复杂背景下的物体检测问题，引入了PIoU损失函数以提高检测精度。