Small objects detection is a challenging task in computer vision due to itslimited resolution and information. In order to solve this problem, themajority of existing methods sacrifice speed for improvement in accuracy. Inthis paper, we aim to detect small objects at a fast speed, using the bestobject detector Single Shot Multibox Detector (SSD) with respect toaccuracy-vs-speed trade-off as base architecture. We propose a multi-levelfeature fusion method for introducing contextual information in SSD, in orderto improve the accuracy for small objects. In detailed fusion operation, wedesign two feature fusion modules, concatenation module and element-sum module,different in the way of adding contextual information. Experimental resultsshow that these two fusion modules obtain higher mAP on PASCALVOC2007 thanbaseline SSD by 1.6 and 1.7 points respectively, especially with 2-3 pointsimprovement on some smallobjects categories. The testing speed of them is 43and 40 FPS respectively, superior to the state of the art Deconvolutionalsingle shot detector (DSSD) by 29.4 and 26.4 FPS. Keywords: small objectdetection, feature fusion, real-time, single shot multi-box detector
展开▼
机译：由于其有限的分辨率和信息，小物体检测在计算机视觉中是一项具有挑战性的任务。为了解决这个问题，现有方法的大多数牺牲了速度以提高精度。在本文中，我们的目标是使用BestObject Detector Single Shot Multibox Detector（SSD）来以较快的速度检测较小的物体，并以精度与速度之间的权衡为基础。为了提高小对象的准确性，我们提出了一种在SSD中引入上下文信息的多层次特征融合方法。在详细的融合操作中，我们设计了两个特征融合模块，即串联模块和元素求和模块，不同之处在于添加上下文信息的方式不同。实验结果表明，这两种融合模块在PASCALVOC2007上获得的mAP分别比基准SSD高1.6和1.7点，特别是在某些小物体类别上提高了2-3点。它们的测试速度分别为43和40 FPS，比最先进的反卷积单发检测器（DSSD）高29.4和26.4 FPS。关键字：小物体检测，特征融合，实时，单发多盒检测器
展开▼
