Person search: Joint Detection and Identification Feature Learning for Person Search

(结合了detection 和 re-id问题)

OIM提出的背景：

1/(triplet loss) not efficient as only several data samples are compared at each time.

(softmax) but finding efficient sampling strategies becomes much more difficult as N increases

2/learning to classify identities with the Softmax loss function [35], which effectively compares all the samples at the same time.

3/We thus exploit a circular queue to store their features also for comparison.

循环队列储存labeled identities 和一个negatives identities队列

网络结构：

前面是resnet50和Faster-rcnn 的东西

识别是RoI-Pooling layer [9] to pool a 1024 × 14 × 14 region from the stem feature maps for each proposal. Then they are passed through the rest conv4 4 to conv5 3 of the ResNet-50, followed by a global average pooling layer to summarize into a 2048 dimensional feature vector.

OIM

为什么不用softmax：

during each SGD iteration we only have positive samples from tens of classes. The classifier matrix suffers from large variance of gradients and thus cannot be learned effectively, even with proper pre-training and high momentum. Second, we cannot exploit the unlabeled identities with Softmax loss.

To overcome this problem, we can approximate the denominators by sub-sampling the labeled and unlabeled identities,