search

PengBoXiangShang / SiamRPN_plus_plus_PyTorch

SiamRPN, SiamRPN++, unofficial implementation of "SiamRPN++" (CVPR2019), multi-GPUs, LMDB.
483 stars 90 forks source link

worse result for test auc in otb100 #9

Open bhr001 opened 5 years ago

bhr001 commented 5 years ago

thx for this great job at first！！！ i trained this code ONLY in ILSVRC2015_finetune,not like mult-trained(eg: trained in COCO,YoutubebBB，ILSVRC2015) in paper,and i got the auc:0.50 for epoch 20 in OTB100 testset.It's much worse than the results in the paper（0.69）. so Reasons for this phenomenon should be Too little training data？（only in ILSVRC2015）

SampsonKwan commented 5 years ago

So do i. I trained on ILSVRC2015, and evaluated on VOT2018 and only got ACC: 0.47 (epoch 20), the results reported in the paper is 0.6. @PengBoXiangShang Would you mind providing the pretrained checkpoint or evaluation code? Gratefully!

klyjb commented 5 years ago

@bhr001 @SampsonKwan How do you test? I wonder where the test code is. Thank you

bhr001 commented 5 years ago

@klyjb you can get test code from these two below . https://github.com/HelloRicky123/Siamese-RPN https://github.com/STVIR/pysot

TeslaHua commented 5 years ago

@bhr001 Hello, I have some questions about the recurrence of SiamRPN++. Please contact me QQ: 975380547

bhr001 commented 5 years ago

@TeslaHua whats matter with your code？

TeslaHua commented 5 years ago

@bhr001 Have you improved the effect of running the model reproduced by this author? Have you found out that this model is different from the SiamRPN++ model in the paper?

bhr001 commented 5 years ago

你说的不同在哪里？我觉得模型一样的，结果不一样是因为paper里在多数据集训练，但现在这个作者的代码只在一个数据集训练，所以泛化能力远不够，分不高

------------------ 原始邮件 ------------------ 发件人: "华子"notifications@github.com; 发送时间: 2019年5月24日(星期五) 下午4:39 收件人: "PengBoXiangShang/SiamRPN_plus_plus_PyTorch"SiamRPN_plus_plus_PyTorch@noreply.github.com; 抄送: "白浩然"522184271@qq.com; "Mention"mention@noreply.github.com; 主题: Re: [PengBoXiangShang/SiamRPN_plus_plus_PyTorch] worse result fortest auc in otb100 (#9)

@bhr001 Have you improved the effect of running the model reproduced by this author? Have you found out that this model is different from the SiamRPN++ model in the paper?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

TeslaHua commented 5 years ago

@SampsonKwan How to test on VOT?