OHEM源码部分

最新推荐文章于 2022-03-14 17:26:56 发布

cycyco

最新推荐文章于 2022-03-14 17:26:56 发布

阅读量195

点赞数 1

分类专栏：深度学习文章标签： deep learning 深度学习 github python

本文链接：https://blog.csdn.net/qq_38591130/article/details/110390940

版权

深度学习专栏收录该内容

12 篇文章 0 订阅

订阅专栏

该博客介绍了一种在线困难样本挖掘（OHEM）方法在基于Fast R-CNN的目标检测框架中如何使用。通过NMS进行去重和多样性选择，挑选出损失最大的样本，用于构建训练批次，从而优化模型训练。源码链接提供了具体的实现细节。

摘要由CSDN通过智能技术生成

在线难例挖掘OHEM
论文Training Region-based Object Detectors with Online Hard Example Mining
源码链接https://github.com/abhi2610/ohem
基于Fast R-CNN，框架基于Caffe

OHEM训练样本生成的部分在/ohem/blob/master/lib/roi_data_layer/minibatch.py里的 get_ohem_minibatch函数

def get_ohem_minibatch(loss, rois, labels, bbox_targets=None,
                       bbox_inside_weights=None, bbox_outside_weights=None):
    """Given rois and their loss, construct a minibatch using OHEM."""
    loss = np.array(loss)

    if cfg.TRAIN.OHEM_USE_NMS:
        # Do NMS using loss for de-dup and diversity
        keep_inds = []
        nms_thresh = cfg.TRAIN.OHEM_NMS_THRESH
        source_img_ids = [roi[0] for roi in rois]
        for img_id in np.unique(source_img_ids):
            for label in np.unique(labels):
                sel_indx = np.where(np.logical_and(labels == label, \
                                    source_img_ids == img_id))[0]
                if not len(sel_indx):
                    continue
                boxes = np.concatenate((rois[sel_indx, 1:],
                        loss[sel_indx][:,np.newaxis]), axis=1).astype(np.float32)
                keep_inds.extend(sel_indx[nms(boxes, nms_thresh)])

        hard_keep_inds = select_hard_examples(loss[keep_inds])
        hard_inds = np.array(keep_inds)[hard_keep_inds]
    else:
        hard_inds = select_hard_examples(loss)

    blobs = {'rois_hard': rois[hard_inds, :].copy(),
             'labels_hard': labels[hard_inds].copy()}
    if bbox_targets is not None:
        assert cfg.TRAIN.BBOX_REG
        blobs['bbox_targets_hard'] = bbox_targets[hard_inds, :].copy()
        blobs['bbox_inside_weights_hard'] = bbox_inside_weights[hard_inds, :].copy()
        blobs['bbox_outside_weights_hard'] = bbox_outside_weights[hard_inds, :].copy()

    return blobs