arbitularov/siamrpn-pytorch

Implementation SiamRPN on PyTorch with GOT-10k dataset

I got on OTB2015 result 8.41 and 0.625 without RPN. In the SiamFusion project

SiamRPN-PyTorch

Implementation SiamRPN on PyTorch with GOT-10k dataset

How to run Training

1. Download the GOT-10k dataset in http://got-10k.aitestunion.com/downloads
2. Run the train_siamrpn.py script:

cd train

python3 train_siamrpn.py --train_path=/path/to/dataset/GOT-10k/train

How to run Tracking

[Coming Soon]

pip install

pip3 install shapely

How to fix GOT-10k dataset

1. First you need to delete four videos:

GOT-10k_Train_008628
GOT-10k_Train_008630
GOT-10k_Train_009058  
GOT-10k_Train_009059

Because they are ymin and xmin is greater than the size of the image.

2. Run the fixed.py script:

python3 fixed.py --dataset_path=/path/to/dataset/GOT-10k/train

After you have new_file.txt file. In this file a lot of information about where the error.

You do not need to change anything yourself, the fixed.py script will do it for you.

My contacts

E-mail: [email protected]

WeChat: tularov_arbi

Authors

• Bo Li - paper - Siamese-RPN
• De jiasong - code - Siamese-RPN-pytorch
• Makalo - code - Siamese-RPN-tensorflow

Citation

Paper: @InProceedings{Li_2018_CVPR,
author = {Li, Bo and Yan, Junjie and Wu, Wei and Zhu, Zheng and Hu, Xiaolin},
title = {High Performance Visual Tracking With Siamese Region Proposal Network},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}

siamrpn-pytorch's Issues

no module named 'got10k'什么情况？

运行python3 fixed.py --dataset_path=/path/to/dataset/GOT-10k/train
出现上述错误

transfer learning strategy

Hi.
Can you please share some transfer learning strategy to start with for Siamese networks
without training from scratch?
thank you

test problem

Can you offer your pretrained alexnet model?

I will be appreciated if you can offer that because I can not use the model offered by pytorch model zoo.

why do you use the group conv in the RPN which groups are batch size

F.conv2d(conv_scores, score_filters, groups=N)

train

When training with the first 500 sequences of GOT-10K, the following error is reported:
You are using "default" experiment, my advice to you is: Copy "default" change folder name and change settings in file "parameters.json"
Traceback (most recent call last):
File "train_siamrpn.py", line 201, in
main()
File "train_siamrpn.py", line 46, in main
seq_dataset = GOT10k(root_dir, subset='train')
File "/usr/local/lib/python3.6/dist-packages/got10k/datasets/got10k.py", line 32, in init
self._check_integrity(root_dir, subset)
File "/usr/local/lib/python3.6/dist-packages/got10k/datasets/got10k.py", line 92, in _check_integrity
raise Exception('Dataset not found or corrupted.')
Exception: Dataset not found or corrupted.

How to run tracking？

Thank you very much for your work. But. How to run tracking?
Only load models in run_tracking.py.
USE tracker.updata?

About README.MD

Can you update the README.MD file?

During training, I encountered this problem.RuntimeWarning: invalid value encountered in log target_w = np.log(gt_w / anchor_w)

the error during training

File "/home/SiamRPN-PyTorch/train/data.py", line 136, in open
img_mean_d )
File "/home/SiamRPN-PyTorch/train/data.py", line 192, in get_instance_image
instance_img, scale_x = self.crop_and_pad(img, cx, cy, size_x, s_x, img_mean)
File "/home/SiamRPN-PyTorch/train/data.py", line 215, in crop_and_pad
xmin = int(self.round_up(xmin + left))
ValueError: cannot convert float NaN to integer

how can i fix it？

How's the performance？

How‘s the tracking result using GOT-10k as training set compared with the large volume of YouTube-BB?

Why is the value of the loss function so large

the error during running tracking

Different from the question from " the error during training". I found that the training is OK, but when running tracking, the error always occur. The message of error is Value Error: cannot convert float NaN to integer. I think this is not the caused by unconverged training.

[Is it ok to train with sequence length ==1 ]

Hi everyone,
I wonder is it make sense to train the model with the sequence length==1.
I ask that because I have the custom dataset for training people detector which do not have any sequence information. i.e I just have the label x,y,w,h, classid for each image.

And I want to train a other model that is like a supplement for the people detector model in our software like when the people detector fail to detect people at the current frame, we can use the SiamRPN to infer the bboxes that likely having object in the current frames based on the detected boxes by people detector in the previous frames
Is it make sense and is there any problem to train the model with sequencce length==1?
What do you guys think?
Thank you

Where is the init_weights?

训练数据集

你在got-10k 训练模型嘛？效果比作者在imagenet和youtubb数据集好是吗？

How to test the tracker on OTB or VOT datasets?

Thanks.

Thanks for you job.could you tell me the accuracy tested on OTB or vot datasets?