Yixiao Ge

> @jackie840129 @deep0learning > I don't have much time for re-arranging the pose generation scripts currently. > But you could refer to the original (but ugly) script here: https://github.com/yxgeee/pytorch_Realtime_Multi-Person_Pose_Estimation/blob/generate/gen.sh I...

All the testing configs are following those used in SARE and NetVLAD papers. Please try to find more details in their papers, e.g., the image size. Also note that the...

@Guan2014 Please follow the setup in https://github.com/Relja/netvlad/blob/master/demoRetrieval.m. For example, use PIL.Image.BICUBIC for resizing.

Could you please provide more details? I have re-tested the model and achieved the same results as I reported. Did you test on the test split of Pitts-250k?

没有写入文件夹的权限吧，你可以把logs给chmod -R 777一下

Yes, the batch size would affect the final performance, since the training of batch norm layers in the network is critical for optimal performance.

It shows that the keys of pre-trained weights cannot match the state_dict of the model. Please provide your detailed steps. I cannot localize the issue given such a simple description.

I used `iters=400`. p.s. `iters=800` is only adopted when msmt17 or vehicle datasets are used as the target-domain dataset.

> Thanks for this special project. > > I trained my own dataset, finished all stage. But i want to give a specific source and target image, and generate an...

I guess the synthetic vehicle dataset (VehicleX) for 2021 AI City Challenge is the same as that for 2020 Challenge. If so, you can require the dataset directly from the...