search

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

Visualize #9

Open aswa123 opened 4 years ago

aswa123 commented 4 years ago

I run Vis.py but there is no output i get, how can I Visualize same as demo video?

mks0601 commented 4 years ago
  1. Read https://github.com/facebookresearch/InterHand2.6M#test
  2. Set vis=True in here and here.
  3. Run test.
liwenssss commented 4 years ago

When I use your visualize demo to show the 3d joints, I find the position of the right hand and left hand is opposite, how to set it to show the position in the right position?

mks0601 commented 4 years ago

Hmm could you give me some examples of the opposite results?

liwenssss commented 4 years ago

Hmm could you give me some examples of the opposite results?

the rgb image is like this: image

but the corresponding visualized 3d joint is like this: image

mks0601 commented 4 years ago

This is because I subtract the root joint position at here for the evaluation.

The code makes the two hand 3D coordinates have the same (0,0,0) root joint (wrist) position. For the visualization, you may want to visualize the 3D hand pose before subtracting the root joint position.

chenyanyin commented 3 years ago

Hmm could you give me some examples of the opposite results?

the rgb image is like this: image

but the corresponding visualized 3d joint is like this: image

have you got bbox and abs_depth by rootnet? in other words, i have to inference rootNet and InterNet for get hand-3D results when i test one image not STB/RHD/InterHand2.6M datasets, the image just got for my cam. @liwenssss @mks0601 Thanks!

mks0601 commented 3 years ago

I used GT bbox for the training and testing, as described in the paper. For the RootNet, actually, you can just visualize pred_joint_coord_img right after here without adding the results of RootNet (abs_depth). x and y coordinates of pred_joint_coord_img are in image space, and z coordinate is in the voxelized heatmap space, not mm. However, it will look very similar, so you can just use it.