search

researchmm / SiamDW

[CVPR'19 Oral] Deeper and Wider Siamese Networks for Real-Time Visual Tracking
http://openaccess.thecvf.com/content_CVPR_2019/html/Zhang_Deeper_and_Wider_Siamese_Networks_for_Real-Time_Visual_Tracking_CVPR_2019_paper.html
MIT License
751 stars 180 forks source link

About the groundtruth in GOT10K's json. #47

Closed kasuo511 closed 5 years ago

kasuo511 commented 5 years ago

Thanks for sharing your code. I noticed that the groundtruth in GOT10K's json file is much bigger than the resolution of GOT10K's train pictures. The resolution is 127127 and 255255, but the gt is more like [344, 223, 776, 1002]. I have no idea how it comes, can you please explain it?

JudasDie commented 5 years ago

Thanks for sharing your code. I noticed that the groundtruth in GOT10K's json file is much bigger than the resolution of GOT10K's train pictures. The resolution is 127_127 and 255_255, but the gt is more like [344, 223, 776, 1002]. I have no idea how it comes, can you please explain it?

Thanks for your interest. [344, 223, 776, 1002] refers to the gt box on the original image instead of the input of Siamese net.

kasuo511 commented 5 years ago

Then how can I get to know the original image's size? They do not seem to be in the dataset folder or in json file, but the training code runs correctly.

JudasDie commented 5 years ago

Then how can I get to know the original image's size? They do not seem to be in the dataset folder or in json file, but the training code runs correct There is no need to know original image's size to train SiamNet, since all objects are centered at the pairs.

kasuo511 commented 5 years ago

So whats the point of these functions:_get_image_anno(self, video, track, frame): and def _get_pairs(self, index): code from module.py. It returns every image's path and annotation. What is it for?

JudasDie commented 5 years ago

So whats the point of these functions:_get_image_anno(self, video, track, frame): and def _get_pairs(self, index): code from module.py. It returns every image's path and annotation. What is it for?

The width and height of an object are utilized to calculate delta for RPN training.

kasuo511 commented 5 years ago

Thank you for your patience. At last, I would like to ask some pretrained models. Could you please provide the CIResNet19_RPN and CIResNet16_RPN 's pretrained models?

JudasDie commented 5 years ago

Thank you for your patience. At last, I would like to ask some pretrained models. Could you please provide the CIResNet19_RPN and CIResNet16_RPN 's pretrained models?

You're so welcome. They share same pretrained models with CIRes22.

kasuo511 commented 5 years ago

Or do you mind sharing the trained models for CIResNet19 and CIResNet16 on RPN, I really interested in several different models that's discussed in your paper.

JudasDie commented 5 years ago

Or do you mind sharing the trained models for CIResNet19 and CIResNet16 on RPN, I really interested in several different models that's discussed in your paper.

Would you mind leave you email? I will send to you later.

kasuo511 commented 5 years ago

That would be nice, please email to tianpeng_liu@163.com. Thanks in advance!

JudasDie commented 5 years ago

That would be nice, please email to tianpeng_liu@163.com. Thanks in advance!

You are welcome.