zhanxinrui/hdn

Homography Decomposition Networks for Planar Object Tracking

hdn's Issues

Fine Tuning The Model

I am trying to fine-tune the HDN model on the POT dataset. I have preprocessed the POT dataset as instructed. However, the training loss seems jumpy as it fluctuates from 1 to 600. I tried to debug the training flow and found that it happens when we feed the supervised training samples to the model. Also, the loss for the HMNET model comes out to be nan. Even after training the model for 1 epoch, the model performance decreases too much. Is there a bug in the training pipeline? Please assist.

What is the extract code of POT280 dataset?

I do not found the extract code of POT280 dataset. Could you tell me.? Thanks.

How to decrease the inference time?

Is there any way to decrease the inference time by arranging the network? Or maybe another way to do it...

Build error

Hi
while building I am getting this error

toolkit/utils/region.c: In function ‘pyx_pw_7toolkit_5utils_6region_7Polygon_5__str’:
toolkit/utils/region.c:4498:70: warning: ‘__pyx_v_i’ may be used uninitialized in this function [-Wmaybe-uninitialized]
4498 | __pyx_t_10 = PyFloat_FromDouble((__pyx_v_self->_c_region_polygon->x[__pyx_v_i])); if (unlikely(!__pyx_t_10)) __PYX_ERR(0, 141, __pyx_L1_error)
| ^
toolkit/utils/region.c:4364:8: note: ‘__pyx_v_i’ was declared here
4364 | long __pyx_v_i;

Can anyone address why this happening?

About Training

I tried two different training methods.
one is training without pertained model, but loss is like this and do not decrease

epoch: 11 481/9375 lr: 0.00241 cls_loss_sup: 0.089103 loc_loss_sup: 0.042983 cls_loss_lp_sup: 0.000710 loc_loss_lp_sup: 0.006392 sim_cent_loss: 0.139188 sim_loss: 0.010180 cor_loss: 25.370169 cor_pos_loss: 25.370132 cor_neg_loss: 0.000036 homo_unsup_loss: 0.017426 total_loss: 15.785341
epoch: 11 511/9375 lr: 0.00255 cls_loss_sup: 0.083010 loc_loss_sup: 0.036742 cls_loss_lp_sup: 0.000312 loc_loss_lp_sup: 0.005055 sim_cent_loss: 0.125119 sim_loss: 0.009541 cor_loss: 29.103842 cor_pos_loss: 29.103813 cor_neg_loss: 0.000029 homo_unsup_loss: 0.018057 total_loss: 14.545660
epoch: 11 541/9375 lr: 0.00271 cls_loss_sup: 0.110210 loc_loss_sup: 0.056279 cls_loss_lp_sup: 0.000471 loc_loss_lp_sup: 0.010177 sim_cent_loss: 0.177137 sim_loss: 0.012793 cor_loss: 24.089970 cor_pos_loss: 24.089928 cor_neg_loss: 0.000042 homo_unsup_loss: 0.013832 total_loss: 19.386679
epoch: 11 571/9375 lr: 0.00285 cls_loss_sup: 0.097187 loc_loss_sup: 0.030674 cls_loss_lp_sup: 0.006395 loc_loss_lp_sup: 0.005135 sim_cent_loss: 0.139392 sim_loss: 0.008259 cor_loss: 22.408772 cor_pos_loss: 22.408726 cor_neg_loss: 0.000046 homo_unsup_loss: 0.011788 total_loss: 16.211657
epoch: 11 601/9375 lr: 0.00300 cls_loss_sup: 0.094609 loc_loss_sup: 0.069493 cls_loss_lp_sup: 0.001097 loc_loss_lp_sup: 0.006270 sim_cent_loss: 0.171469 sim_loss: 0.007646 cor_loss: 28.108177 cor_pos_loss: 28.108162 cor_neg_loss: 0.000015 homo_unsup_loss: 0.011420 total_loss: 19.361439
epoch: 11 631/9375 lr: 0.00316 cls_loss_sup: 0.089650 loc_loss_sup: 0.033506 cls_loss_lp_sup: 0.004167 loc_loss_lp_sup: 0.005310 sim_cent_loss: 0.132633 sim_loss: 0.008574 cor_loss: 25.780205 cor_pos_loss: 25.780182 cor_neg_loss: 0.000022 homo_unsup_loss: 0.019004 total_loss: 15.041709
epoch: 11 661/9375 lr: 0.00331 cls_loss_sup: 0.097170 loc_loss_sup: 0.057477 cls_loss_lp_sup: 0.001001 loc_loss_lp_sup: 0.005108 sim_cent_loss: 0.160755 sim_loss: 0.005386 cor_loss: 28.488239 cor_pos_loss: 28.488205 cor_neg_loss: 0.000035 homo_unsup_loss: 0.011990 total_loss: 18.341190

and I use the model to inference and got the bad results.
can you tell me the total_loss when you train the model?

another one is train model based on the pretrained model you provided and I also get the bad result.

Looking forward to your reply,please.

best model是如何训练的

Xinrui Zhang您好，我想请教一下，你们的release出来的best model是如何训练的？

best model指hdn-simi-sup-hm-unsup.pth

我注意到论文中的Training Details一节中写道你们用COCO14和GOT10k做了增强，请问具体的训练过程是什么样的？

1. 先用GOT-10k和COCO14做pretrain，再在POT上做finetune
2. 直接用GOT-1k、COCO14和POT做一次train，得到best model

我现在无法复现出和best model一样的结果，可能是我的训练方法或config设置不同。如果方便的话，能否一并提供config.yaml？

can not download Demo video

Dear Master,

I want to check Demo video through attached link (https://www.aliyundrive.com/s/ycDqPLz5e3Z/folder/61bef2d66bbb3e8803544251adf525354f3240bd)
But, I don't know how to download it....
Could you please share Demo video through Google Drive?

Thanks,
kwanghyuk

access code

Hi, it required the access code of this link:https://pan.baidu.com/s/1boKIoXGFOWZ-uu9X6WzDCA. Please share it, thanks! @zhanxinrui