The field of object detection has made significant advances riding on thewave of region-based ConvNets, but their training procedure still includes manyheuristics and hyperparameters that are costly to tune. We present a simple yetsurprisingly effective online hard example mining (OHEM) algorithm for trainingregion-based ConvNet detectors. Our motivation is the same as it has alwaysbeen -- detection datasets contain an overwhelming number of easy examples anda small number of hard examples. Automatic selection of these hard examples canmake training more effective and efficient. OHEM is a simple and intuitivealgorithm that eliminates several heuristics and hyperparameters in common use.But more importantly, it yields consistent and significant boosts in detectionperformance on benchmarks like PASCAL VOC 2007 and 2012. Its effectivenessincreases as datasets become larger and more difficult, as demonstrated by theresults on the MS COCO dataset. Moreover, combined with complementary advancesin the field, OHEM leads to state-of-the-art results of 78.9% and 76.3% mAP onPASCAL VOC 2007 and 2012 respectively.
展开▼
机译：在基于区域的ConvNets的浪潮中，对象检测领域取得了长足的进步，但是它们的训练过程仍然包括许多启发式方法和超参数，它们的调整成本很高。我们提出了一个简单但令人惊讶的有效在线硬示例挖掘（OHEM）算法，用于训练基于区域的ConvNet检测器。我们的动机与往常一样-检测数据集包含大量简单示例和少量困难示例。自动选择这些困难的例子可以使培训更加有效。 OHEM是一种简单直观的算法，它消除了一些常用的试探法和超参数。但更重要的是，它在PASCAL VOC 2007和2012等基准上的检测性能始终如一且显着提高。随着数据集变得越来越大，越来越困难，其有效性也随之提高。 MS COCO数据集上的结果。此外，结合该领域的互补进展，OHEM在PASCAL VOC 2007和2012上分别获得了78.9％和76.3％的mAP的最新结果。
展开▼
