facebookresearch / InterHand2.6M

Official PyTorch implementation of "InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image", ECCV 2020
Other
687 stars 91 forks source link

How's the data organized differently from v0.0? #68

Closed zc-alexfan closed 3 years ago

zc-alexfan commented 3 years ago

I have been using v0.0 and recently need to adapt to the new v1.0 version. The naming convention for the annotation files seems to have changed a lot. My code is expecting the different splits for train, val, test in terms of all, machine_annot, human_annot. However, the current annotation look like the following:

train ls
InterHand2.6M_train_MANO_NeuralAnnot.json  InterHand2.6M_train_camera.json  InterHand2.6M_train_data.json  InterHand2.6M_train_joint_3d.json

val ls
InterHand2.6M_val_MANO_NeuralAnnot.json  InterHand2.6M_val_camera.json  InterHand2.6M_val_data.json  InterHand2.6M_val_joint_3d.json

test ls
InterHand2.6M_test_MANO_NeuralAnnot.json  InterHand2.6M_test_camera.json  InterHand2.6M_test_data.json  InterHand2.6M_test_joint_3d.json

Are they all the all split? Or how do I find the machine_annot, and human_annot?

mks0601 commented 3 years ago

Hi,

As many people got confused between all/machine_annot/human_annot and train/test/val, I've decided to provide only all split. So, the question to your answer is : yes. all json files are all split. Do you want machine_annot and human_annot?

zc-alexfan commented 3 years ago

I see. Do you also have machine_annot and human_annot?

Another question: for 5 FPS, the v1.0 version differs from v0.0 in terms of number of images? I ran my experiments for my submission using v0.0. I guess I need to rerun them on v1.0.

mks0601 commented 3 years ago

I can provide annotation ids of machine_annot and human_annot so that you can use them from all. In both 5 and 30fps, v1.0 provides full images as written in here. v0.0 has only half of full images.

zc-alexfan commented 3 years ago

That works. Thanks

mks0601 commented 3 years ago

Here are links 5fps: https://drive.google.com/drive/folders/1iDI_wRkngB374IksPbEB_OeoO1CJsedV?usp=sharing 30fps: https://drive.google.com/drive/folders/1bCqioHhiXiB3Mva1hbBENeY4vZjtfW62?usp=sharing

wbhu commented 3 years ago

Hi @mks0601,

Thanks for your great work. I also need the split for machine_annot and human_annot. But I only found the 'train' and 'test' splits for machine_annot and human_annot, while the original annot has 'train', 'val', and 'test' splits. What is the correspondence?

Thanks

mks0601 commented 3 years ago

Hi,

Actually, val is defined only for machine_annot. The provided val in here and here are from machine_annot.

wbhu commented 3 years ago

got it, thanks