nileshkulkarni / acsm Goto Github PK

Hi Nilesh,

Is it possible to share the annotation and data used for the cow and sheep categories used (just like you did for the horse category)?

Thanks a lot in advance!

Best,
Anastasis

Hi,

Thanks for your excellent work and contribution to the community!

I noticed the models for cow and sheep are missing (the respective folders in /cachedir/models are empty for these 2 animals). Is it possible to share the content of those folders?

Thanks a lot in advance!

Best,
Anastasis

hello. Several years have passed since your paper was published, but I am amazed at the thought that your paper is still valid. As you can see from my profile picture, I am very interested in cats and would like to use cats to check if your algorithm is valid. Among them, detailed information is needed on how to create the mean_shape.mat file. Can you help me?

thank you

Thank you for your brilliant work!
I have read your excellent paper -ACSM and would like to ask some questions about some details in parameterizing part transformations in articulation procedure.

As you showed in the supplemental,
"We parameterize part transforms as an axis, angle representation." and
"Every part’s axis serves as a bias in the network that is learned and is same across the whole category."

I would like to ask, what is the axis and angle stands for?
From my perspective axis = (x0,y0,z0) and angle is represented as quaternion, is it right?

How to depict the vertices after transformation.
p_new = R * (p_old - bias) + bias (1) (R is rotation matrix)
Did you do the transformation in this way?
So each part of bias is assigned (like pick some fix point) or need to learned by deep network?

How to learn the angle?
Like using ResNet + FC similar to the camera prediction since they are both representation for angle?

Thank you for checking my problem and I would like to receive your reply.

As you can see, I have group the obj with Meshmixer and get the same vertices and faces as the original obj. But it can't be grouped. The horse_part.obj you gave me can be grouped and has 708 vertices. So I think it may due to this difference. Could you tell me more about group?

Best,
Peisen Lin

Hi, thanks for the great publication. A quick question on using a-csm on other custom parametric models. For inferencing on categories not present in CUBS, is there a way to use the given template models and annotations?

Thank you for your excellent work!

I have read your paper and have questions: how did you annotate vertices of 3d-shape, e.g. horse?

Thank you for your excellent work!

I have run your acsm code and found the PCK under reprojection for horse is much smaller than you show in your paper. The PCK transfer is fine.

as you can see in the following picture, the predict key points are not so close to the ground truth.


I have run this to test:
python -m acsm.benchmark.pascal.kp_project --name=acsm_horse_8parts --category=horse --parts_file=acsm/part_files/horse.txt --use_html --dl_out_pascal=True --dl_out_imnet=False --split=val --num_train_epoch=400 --num_hypo_cams=8 --env_name=acsm_horse_8parts_pck_val --multiple_cam=True --visuals_freq=5 --visualize=True --n_data_workers=4 --scale_bias=0.75 --resnet_style_decoder=True --resnet_blocks=4
Have I done something wrong?