Hands2019 chanllenge code about a2j HOT 7 OPEN

Hello, our hands2019 challenge code is basically the same with hands2017, you can refer to our paper for more implementation details.

from a2j.

Hello, our hands2019 challenge code is basically the same with hands2017, you can refer to our paper for more implementation details.

Looks like hands2019 dataset is different in size: In your paper you mention there are 295K test images for hands2017. However, in hands2019, there are 125K, 25K and 6K test images for Task 1, 2, 3 respectively.
Could you please explain the structure of /data/zhangboshen/CODE/Anchor_Pose_fpn/data/Hands2017/frame/images/?

from a2j.

Looks like hands2019 dataset is different in size: In your paper you mention there are 295K test images for hands2017. However, in hands2019, there are 125K, 25K and 6K test images for Task 1, 2, 3 respectively.
Could you please explain the structure of /data/zhangboshen/CODE/Anchor_Pose_fpn/data/Hands2017/frame/images/?

This file folder /data/zhangboshen/CODE/Anchor_Pose_fpn/data/Hands2017/frame/images/ is the original testset of HANDS17 challenge.
And yes, HANDS19 and HANDS17 are different in size, but after training phase. Just simply replace testset directory, this whole code thing is the same with HANDS17.

from a2j.

I understand. Could you share how to generate center_file for HANDS19, and what MEAN and STD represent so we can compute them for HANDS19 as well? Thanks.

from a2j.

We use center_file generated by this repo:https://github.com/mks0601/V2V-PoseNet_RELEASE,
And in HANDS19, we only use bndbox to crop the image, without using center_file.
MEAN and STD are computed by averaging the mean and std of whole training set (after croping images).

from a2j.

I see, thanks. Could you share the code for HANDS19?

from a2j.

Thanks for the amazing code Boshen! I am also very interested in the training scripts for HANDS2019, especially for Task2 (hand-object interaction). It'll be very helpful if you'd provide some code for those as well. Thanks!

from a2j.