dragonbook / v2v-posenet-pytorch Goto Github PK

Hi, Thanks for the implementation of V2V-Posenet in Pytorch. It's really helpful in understanding the authors' work. I'm thinking of using your code in a project and wonder if it's open-source. If it is, would it be possible to add a license?

Thanks again.

Your code works fine in the MSRA dataset. I changed the code to the NYU dataset. The precision is not up to the torch7 version. I used your parameter settings with an accuracy of 13mm, using the author's parameter settings. 10mm, what do you think is the reason?

Why I get an image of a hand gesture standing on its head(all the results are the same!):

Do you consider providing training code as it provide in https://github.com/mks0601/V2V-PoseNet_RELEASE

Thank you for sharing your hard work. It has been so helpful to me.

In your implementation, I could find code for MSRA dataset only.

Are ITOP dataset, which is included in the V2V-PoseNet author's torch 7 implementation, also possible?

If it is not, can there be any critical differences I should keep in mind while implementing ITOP by myself?

Thank you so much and have a good day.

Do you have pretrained model?

the sample numble of msra_center provided in your code dont equal msra dataset. one(msra_center) is 76391 and another(msra dataset) is 76375.

draw_DB.m file is missing can u give me the file?
After what are the clear steps to follow I'm having the epoch*.pth files where to give to see the results of image in png format-Paper_result_NYU.png

Thanks for your sharing.

I have changed your msra_hand.py to nyu_hand.py fitting NYU dataset and keep all other parts of your code unchanged. However, the result of NYU dataset is bad. The all mean error is 180mm after training. (epoch=14, batch_size=64, nGpus=4, optimizer =Adam, cubicSize=250, keep orignal_size and crop_size the same as yours.
While the results on NYU is about 13mm or less after epoch ensemble with nGPU=1 and batchSz=16.

Here is my question. Why with the increase of batch size the results becoming worse?
Is there any way to solve such a problem?

Thanks again!

Hello,
Thanks for sharing.

In real industrial environment except UYU, ICVL and MSRA datasets, gives a demo for using pretrained model with the depth image from RGBD camera?

If there exists similar Python demos, will be very great and helpful.

Thanks for reply.

Best Regards!

Hi, Thanks a lot for your awesome codebase. Could you please clarify the data loader for hands 2017 dataset?
I could not find that here

Hi,

I've trained your pytorch version model using the MSRA dataset and it works well, and the next step I'd like to use the trained model to run on my own dataset, may I ask how to prepare the centered data like the paper described?

Thanks!

Hi， I have download your pretrained models. But I still have some problems for me to convert my own depth image into the network, and get a good pose/hand estimation. Will you provide some example codes?

Hello and thanks for sharing your work,

Does your code work with pose estimation examples?

Thanks

I am working on a project in which i have to estimate/get the 3d coordinates of the hand from a real-time dataset or a video. I would like to know whether it could get me the coordinates from a video or not. If it could happen then what is the approx fps of the video from which it could extract ?

Could you provide the code of hand voxel visualization?

bearpaw/pytorch-pose#37 just like this.