• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>International Conference on Pattern Recognition >Multiple-step Sampling for Dense Object Detection and Counting
【24h】

Multiple-step Sampling for Dense Object Detection and Counting

机译：用于密度物体检测和计数的多步采样

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

A multitude of similar or even identical objects are positioned closely in dense scenes, which brings about difficulties in object-detecting and object-counting. Since the poor performance of Faster R-CNN, recent works prefer to detect dense objects with the utilization of multi-layer feature maps. Nevertheless, they require complex post-processing to minimize overlap between adjacent bounding boxes, which reduce their detection speed. However, we find that such a multi-layer prediction is not necessary. It is observed that there exists a waste of ground-truth boxes during sampling, causing the lack of positive samples and the final failure of Faster R-CNN training. Motivated by this observation we propose a multiple-step sampling method for anchor sampling. Our method reduces the waste of ground-truth boxes in three steps according to different rules. Besides, we balance the positive and negative samples, and samples at different quality. Our method improves base detector (Faster R-CNN), the detection tests on SKU-110K and CARPK benchmarks indicate that our approach offers a good trade-off between accuracy and speed.
机译：多种相似甚至相同的对象密切地定位在密集的场景中，这在对象检测和对象计数中带来了困难。由于R-CNN的性能较差，最近的作品更倾向于利用多层特征图来检测密集对象。然而，它们需要复杂的后处理，以最小化相邻边界框之间的重叠，这降低了它们的检测速度。但是，我们发现不需要这样的多层预测。观察到采样期间存在浪费地面真理盒，导致缺乏正样品和更快的R-CNN培训的最终失败。通过该观察结果，我们提出了一种用于锚定采样的多步采样方法。我们的方法根据不同的规则减少了三个步骤中的地面真理盒的浪费。此外，我们平衡了正面和阴性样本，并以不同的质量样品。我们的方法改进了基础探测器（更快的R-CNN），SKU-110K和Carpk基准测试的检测测试表明我们的方法在准确性和速度之间提供了良好的权衡。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号