jiawei151 / sgiel_vireid Goto Github PK
View Code? Open in Web Editor NEWOfficial PyTorch Implementation of "Shape-Erased Feature Learning for Visible-Infrared Person Re-Identification" (CVPR'23)
License: MIT License
Official PyTorch Implementation of "Shape-Erased Feature Learning for Visible-Infrared Person Re-Identification" (CVPR'23)
License: MIT License
您好!感谢您精彩的工作。我看到本文的baseline,Rank-1精度大约到了70左右。我去掉您的所有项,训练出来精度为67.99。但是,每个step更新参数后,也执行了EMA更新,但是ema模型最终精度为1%,而且整个训练过程中测试也没啥变化,不知道什么情况。还有就是不加ema 67.99的精度是正常的吗?
论文中有展示HITSZ-VCM上的结果,可是代码不支持在HITSZ-VCM上训练和测试,能提供更多关于HITSZ-VCM的实验细节吗?例如,训练和测试时是video-based还是image-based,video-based的序列长度等信息。
您好,根据您提供的代码和shape的数据集,我在SYSU-MM01数据集上进行了复现。但是我的复现并没有得到论文中的效果。最终跑了200个epoch,最好的效果出现在68个epoch。我的复现结果如下:
==> Preparing Data Loader...
68
current lr 0.001
Epoch:[68][0/695]L:0.0341(0.0341) L2:0.1142(0.1142) sL:0.0230(0.0230) w1:0.5949(0.5949) or:0.0023(0.0023) ML2:3.5368(3.5368) KL:0.8719(0.8719)
Epoch:[68][50/695]L:0.0217(0.0410) L2:0.0868(0.1193) sL:0.0165(0.0280) w1:0.6459(0.5807) or:0.0022(0.0022) ML2:2.8971(3.5661) KL:0.7459(0.8943)
Epoch:[68][100/695]L:0.0179(0.0399) L2:0.0711(0.1145) sL:0.0311(0.0286) w1:0.6091(0.5809) or:0.0023(0.0022) ML2:2.3823(3.4528) KL:0.7855(0.8820)
Epoch:[68][150/695]L:0.0415(0.0398) L2:0.1330(0.1147) sL:0.0267(0.0281) w1:0.6104(0.5807) or:0.0021(0.0022) ML2:3.8870(3.4472) KL:0.8610(0.8742)
Epoch:[68][200/695]L:0.0621(0.0401) L2:0.1400(0.1156) sL:0.0349(0.0285) w1:0.4506(0.5763) or:0.0023(0.0022) ML2:4.5594(3.4788) KL:0.9806(0.8781)
Epoch:[68][250/695]L:0.0520(0.0389) L2:0.1278(0.1129) sL:0.0287(0.0281) w1:0.6022(0.5826) or:0.0024(0.0022) ML2:3.1366(3.4127) KL:0.9168(0.8736)
Epoch:[68][300/695]L:0.0466(0.0392) L2:0.1424(0.1133) sL:0.0282(0.0280) w1:0.5992(0.5830) or:0.0023(0.0022) ML2:4.0995(3.4141) KL:0.9180(0.8747)
Epoch:[68][350/695]L:0.0210(0.0388) L2:0.0708(0.1128) sL:0.0247(0.0279) w1:0.7259(0.5835) or:0.0023(0.0022) ML2:2.3554(3.4080) KL:0.7897(0.8747)
Epoch:[68][400/695]L:0.0293(0.0386) L2:0.0920(0.1121) sL:0.0273(0.0279) w1:0.6834(0.5844) or:0.0022(0.0022) ML2:3.0099(3.3912) KL:0.8538(0.8732)
Epoch:[68][450/695]L:0.0238(0.0390) L2:0.0866(0.1127) sL:0.0356(0.0282) w1:0.7104(0.5832) or:0.0023(0.0022) ML2:2.6720(3.4074) KL:0.8792(0.8757)
Epoch:[68][500/695]L:0.0432(0.0390) L2:0.1183(0.1127) sL:0.0270(0.0281) w1:0.6754(0.5852) or:0.0022(0.0022) ML2:2.9553(3.4083) KL:0.8770(0.8772)
Epoch:[68][550/695]L:0.0225(0.0386) L2:0.0828(0.1121) sL:0.0203(0.0279) w1:0.4978(0.5850) or:0.0023(0.0022) ML2:2.6331(3.3943) KL:0.7263(0.8749)
Epoch:[68][600/695]L:0.0577(0.0386) L2:0.1230(0.1120) sL:0.0318(0.0278) w1:0.4661(0.5841) or:0.0022(0.0022) ML2:3.7425(3.3959) KL:0.9794(0.8749)
Epoch:[68][650/695]L:0.0268(0.0387) L2:0.1098(0.1123) sL:0.0305(0.0278) w1:0.6184(0.5833) or:0.0023(0.0022) ML2:3.7047(3.4044) KL:0.9451(0.8748)
Test Epoch: 68
Extracting Gallery Feature...
Extracting Time: 1.526
Extracting Query Feature...
Extracting Time: 6.547
Evaluation Time: 5.361
Extracting Gallery Feature...
Extracting Time: 1.545
Extracting Query Feature...
Extracting Time: 6.483
Evaluation Time: 5.422
POOL: Rank-1: 67.84% | Rank-5: 88.98% | Rank-10: 94.74%| Rank-20: 98.66%| mAP: 65.55%| mINP: 53.43%
FC: Rank-1: 70.42% | Rank-5: 91.38% | Rank-10: 95.98%| Rank-20: 98.74%| mAP: 67.05%| mINP: 53.69%
Best Epoch [33]
------------------ema eval------------------
POOL: Rank-1: 68.45% | Rank-5: 89.53% | Rank-10: 94.61%| Rank-20: 98.24%| mAP: 66.25%| mINP: 54.22%
FC: Rank-1: 72.31% | Rank-5: 91.48% | Rank-10: 95.69%| Rank-20: 98.40%| mAP: 68.09%| mINP: 54.34%
Best Epoch [68]
我的超参数都是根据代码中的默认设置的。请问这可能是什么问题呢
it seems that you trained your proposed model with the shape data prepared (not generating them when training)
but when i try SCHP model on SYSU, there is a phenomenon that SCHP significantly fails on images with under certain resolution. it outputs totally nothing. how did you solve that? did you manually labeled them or use another model?
and additionally, how did you manage images that have more than 1 person? if more than 1 person exist in an image, SCHP detect all human parts.
您好,请问下您生成的SHAPE 图片是单通道的该怎么处理呢, 直接广播成三个通道吗?
I want to ask how to obtain SYSU_MM01_SHAPE and SYSU_MM01_MASK ?
Could you share your .npy files for the body shape of SYSU-MM01 and RegDB datasets?
我想问一下,data_path1里面是您分享的shape数据集,data_path2文件夹下放的是什么呢?
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.