目标检测之OHEM介绍

论文地址：https://arxiv.org/pdf/1604.03540.pdf

在two-stage检测算法中，RPN阶段会生成大量的检测框，由于很多时候一张图片可能只会有少量几个目标，也就是说绝大部分框是没有目标的，为了减少计算就需要进行sample，一般来说fasterrcnn的sample机制是算框和label的IOU，大于0.7认为是正样本，小于0.3是负样本。但是单纯的random_sample选出来的框不一定是最容易错的框。那么ohem就是这样的一种正负样本策略，通过根据框的loss得到最容易错的框。可以理解为错题集，我们只会把最容易错的题放到错题集。

首先是 negative，即负样本，其次是 hard，说明是困难样本，也可以说是容易将负样本看成正样本的那些样本，例如 RPN框里没有物体，全是背景，这时候分类器很容易正确分类成背景，这个就叫 easy negative；如果 框里有二分之一个物体，标签仍是负样本，这时候分类器就容易把他看成正样本，这时候就是 had negative。hard negative mining 就是多找一些 hard negative 加入负样本集，进行训练。
接下来我们来看看mmdection的ohem实现：

class OHEMSampler(BaseSampler):r"""Online Hard Example Mining Sampler described in `Training Region-basedObject Detectors with Online Hard Example Mining<https://arxiv.org/abs/1604.03540>`_."""def __init__(self,num,pos_fraction,context,neg_pos_ub=-1,add_gt_as_proposals=True,**kwargs):super(OHEMSampler, self).__init__(num, pos_fraction, neg_pos_ub,add_gt_as_proposals)self.context = contextif not hasattr(self.context, 'num_stages'):self.bbox_head = self.context.bbox_headelse:self.bbox_head = self.context.bbox_head[self.context.current_stage]def hard_mining(self, inds, num_expected, bboxes, labels, feats):with torch.no_grad():rois = bbox2roi([bboxes])if not hasattr(self.context, 'num_stages'):bbox_results = self.context._bbox_forward(feats, rois)else:bbox_results = self.context._bbox_forward(self.context.current_stage, feats, rois)cls_score = bbox_results['cls_score']loss = self.bbox_head.loss(cls_score=cls_score,bbox_pred=None,rois=rois,labels=labels,label_weights=cls_score.new_ones(cls_score.size(0)),bbox_targets=None,bbox_weights=None,reduction_override='none')['loss_cls']_, topk_loss_inds = loss.topk(num_expected)return inds[topk_loss_inds]def _sample_pos(self,assign_result,num_expected,bboxes=None,feats=None,**kwargs):"""Sample positive boxes.Args:assign_result (:obj:`AssignResult`): Assigned resultsnum_expected (int): Number of expected positive samplesbboxes (torch.Tensor, optional): Boxes. Defaults to None.feats (list[torch.Tensor], optional): Multi-level features.Defaults to None.Returns:torch.Tensor: Indices  of positive samples"""# Sample some hard positive samplespos_inds = torch.nonzero(assign_result.gt_inds > 0, as_tuple=False)if pos_inds.numel() != 0:pos_inds = pos_inds.squeeze(1)if pos_inds.numel() <= num_expected:return pos_indselse:return self.hard_mining(pos_inds, num_expected, bboxes[pos_inds],assign_result.labels[pos_inds], feats)def _sample_neg(self,assign_result,num_expected,bboxes=None,feats=None,**kwargs):"""Sample negative boxes.Args:assign_result (:obj:`AssignResult`): Assigned resultsnum_expected (int): Number of expected negative samplesbboxes (torch.Tensor, optional): Boxes. Defaults to None.feats (list[torch.Tensor], optional): Multi-level features.Defaults to None.Returns:torch.Tensor: Indices  of negative samples"""# Sample some hard negative samplesneg_inds = torch.nonzero(assign_result.gt_inds == 0, as_tuple=False)if neg_inds.numel() != 0:neg_inds = neg_inds.squeeze(1)if len(neg_inds) <= num_expected:return neg_indselse:neg_labels = assign_result.labels.new_empty(neg_inds.size(0)).fill_(self.bbox_head.num_classes)return self.hard_mining(neg_inds, num_expected, bboxes[neg_inds],neg_labels, feats)

上面代码就是整个ohem的sample过程，整个ohem分为三个函数分别是hard_mining，_sample_pos，_sample_neg，_sample_pos和_sample_neg是获得对应的困难正样本/困难负样本，由hard_mining完成整个sample过程：根据输入的box_list得到对应的bbox_loss的list取最大的256/512个，由于这一批box的loss最大，就可以认为是最难区分的box，这一批bbox就是所谓的
困难正样本/困难负样本。

至此ohem阶段完成，后面就是对候选框的分类和回归，因为ohem阶段得到了容易分错的样本框，所以在后续训练阶段模型会对这些容易分错的框重点关注，有利于困难样本的检测，提升了模型的效果。

实际上提升还是很明显的：