Zhedong Zheng comments

Results 584 comments of


                                            Zhedong Zheng

Training on MSMT

mkl version may be wrong. You may re-install your mkl.

关于视频中的re-id

@wangsun1996 对的，可以优化一下效率。比如一个tracklet就留几张图就行。

Can you try the new ReID data augmentation method? It is reported to be effective in improving the performance of your baseline.

Sure. I will add it in recent days.

Can you try the new ReID data augmentation method? It is reported to be effective in improving the performance of your baseline.

It is added at https://github.com/layumi/Person_reID_baseline_pytorch/blob/master/random_erasing.py#L57

How to identify same person in multicamera using cosine distance metric.

Hi @Basavarajdocketrun When evaluation, you could use the cosine distance as well. Just extract the new feature of the same person in the new camera and do the comparison. By...

PCB structure

Hi @mcjqwer The original paper split the feature into 6 parts, which equals to use [6] adaptive pooling.

请问有考虑三元组损失吗？我在代码中没有找到

You may refer to https://github.com/layumi/Person-reID-triplet-loss

使用swin训练出现的问题

Hi @tjusyy 可能是 timm版本问题。更新timm就行。 Please update your timm and try it again

使用swin训练出现的问题

https://github.com/layumi/Person_reID_baseline_pytorch/blob/master/model.py#L121 要更新到 0.6.0以上。timm 改写法了。 ``` pip install git+https://github.com/rwightman/pytorch-image-models.git ```

Can this project be used in other fields, like clothing retrieval?

Sure. Recently, I also applied it to vehicle retrieval (re-id) competition, yielding the 1st place @clw5180 . You may check more details at https://github.com/layumi/AICIty-reID-2020.