michuanhaohao/keras_reid

Margin Sample Mining Loss: A Deep Learning Based Method for Person Re-identification

Person reid for keras

• Classification: reid_classification.py
• Classification + triplet loss: reid_tripletcls.py (triplet_loss)
• Classification + triplet loss with hard negative mining: reid_tripletcls.py (triplet_hard_loss)
• Classification + margin sample mining loss: reid_tripletcls.py (msml_loss)
• Re-ranking with k-reciprocal Encoding (CVPR2017): re_ranking.py
• Using Pytorch (GPU) to accelerate Re-ranking with k-reciprocal Encoding (CVPR2017): re_ranking_gpu.py
• Pre-trained model: naivehard_more_last.h5 (market1501:81.1% rank-1 accuracy cuhk03:78.8% rank-1 accuracy)

Pre-trained model download

http://pan.baidu.com/s/1bo3gwaV

Note

You should add some functions according to the comments

Citation

@article{xiao2017margin,
  title={Margin Sample Mining Loss: A Deep Learning Based Method for Person Re-identification},
  author={Xiao, Qiqi and Luo, Hao and Zhang, Chi},
  journal={arXiv preprint arXiv:1710.00478},
  year={2017}
}

keras_reid's Issues

how to get data?

where is the 'load_market_img'? I cannot run this code.

data provider

您好，我是一个keras 新手，关于这个数据加载函数可以多给一点提示吗，谢谢!! 如有打扰，还望谅解

How do you normalize datasets?

hello.
How do you normalize datasets?
I tried

img = img.astype("float32")
img /= 127.5
img -= 1.

like it.
What method do you use?

load data

Hi, I am new in keras. Could you please let me know how to load market data? Thanks in advance.

模型的输入问题

你好，我想问一下关于定义的单个输入的模型是怎么输入一个PN*SN的数据进行训练的了？
因为我肯模型定义的输入是一张图片，但是输入返回的train_img,train_label = get_triplet_data(PN)
不应该是一个列表或者数组吗？

期待你的回答，Thanks a lot！

您好, 有个训练问题想问一下

您好, 我最近在测试这个算法.

• 我的数据集大约是1800个Id, 每个Id有6~30+的图, 所以我选择的PN是6.
• 网络结构是Resnet50+Flatten(2048)+FC1800(for cls),
• 对图片的预处理是水平翻转, 因为还没有跑出一个稳定的效果, 就没有增加normalization

出现的问题

• 若将2048的向量作为这张图的feature输出, 我发现在模型训练初期, feature会在100个batch左右变成0, 往后的训练也变成0
• classification的loss从训练开始到很长一段时间都稳定在8左右, triHard_loss在0.6左右(Alpha值)
• 如果只使用分类网络, 那么分类loss也只是稳定在8左右, 并且model.predict出来的结果是同一个Id
• 使用Resnet50的预训练模型, 以及您提供的初始化模型 也都类似以上的结果.
  从数据集的角度上来说应该不是问题关键, 因为在我用cv2在训练时显示了输入网络的图片
  谢谢作者, 希望讨论

where is load_img_data module?

'from load_img_data import get.....' is hard understand;
please tell me the code of load_img_data.

训不出来

表示训不出来

identity_num question

Hello,why does the identity_num need to be 6273 when i load your pre-train weights?

请问一下keras 和tensorflow-gpu的版本是多少呢？

一个小问题

罗博，你好！
看了你写的《基于深度学习的行人重识别研究综述》过来的！写的很赞！
阅读了MLMS losss部分的代码，有一个疑问，即每个batch大小为PN*SN,每次PN个不同的ID，每个ID取SN个样本，最后算损失的时候，我没理解错的话，只得到一个max和一个min，这样是否优点浪费？能否每个ID只跟自己组内算远近，最后每个batch得到PN个max和min呢？

请问使用softmax和triplet loss训练大概需要迭代多少次能收敛？

pretrained model

Hi,
i tried to download the pretrained model but the file that i get is not working (and i dont understand Chinese)
can you please help me to get the h5 file?
thank you

The mlsl loss stays at 0.6

Sorry to bother you, but I finished loaddata function and used mlsl loss func, however, flatten_loss which is mlsl loss was somehow stay at alpha which is set as 0.6 in origin codes.
I think this means positive equals to negative, so that the loss choose to be a larger value which becomes alpha. Is that normal? And the overall loss was not decreasing.
Thanks again!

get_triplet_data

您好，我在看了您关于triplet loss和TriHard loss的keras代码复现，其中负责取triplet loss的batch函数：from load_img_data import get_triplet_data, get_triplet_hard_data，我没有找到相关的库包括注释，向您请教一下，如果打扰了，十分抱歉！

请问一下作者你的预训练模型是怎样训练出来的