gap-lab-cuhk-sz / deepfashion3d Goto Github PK

Hi @kv2000 ,
Superb work and dataset, which in my opinion is a big contribution to this research area.
Meanwhile, I have several questions.
[1] some inconsistency
As mentioned in the abstract, there are 563 garment instances. But according to table 1, the total number is 599.
I suppose you remove some garment instances. But why, e.g., poor SMPL fitting results?

How many random views are used for rendering synthetic images, 5 (section 4.1) or 3 (section 4.3)?
I admit that this does not really matter but somewhat confusing.

[2] Why ignore the shape (i.e., beta) parameter of the SMPL model from the whole pipeline?
The scale of a kid’s cloth is different from that of an adult’s cloth.

[3] The gcn part. In the original Pixel2Mesh paper, the input is always the fixed ellipsoid from which the deformations are.
But in your paper, the input is the varying set of feature lines depending on the category of the cloth (as shown in Fig. 3), right? How does the GCN manage to handle varying set of feature lines?

Thanks very much!

Hi,
thank you for your great work! I can't find RGB image when I using deepFashion3D dataset, where can I get these images?

          So i have two main problems.

1. The default shape is little bigger than the garment.
2. There is some misalignment, the translation/global orient might have some errors
   see attached photo
   

Originally posted by @msverma101 in #11 (comment)

Hi, thanks for you to share the dataset.
I wonder to know how long I can get the pwd of the dataset after submit the form?

Hi! I have a very basic question on loading the data to view. I was trying to view the point cloud, mesh, and the body (from the v2 release) using meshplot but ran into the following issues:

1. not sure which SMPL pkl to use, I don't know where to find the info for whether the body is female, male, or neutral?
2. the packed_pose has field "pose", "scale", and "trans"; I used "pose" and "trans" to initialize the smpl model and then scaled it with the "scale" float but the resulting body mesh and garment mesh do not line up when I load them into the same meshplot plot:

The code I used to create this plot is as follows:

import smplx
import torch
import numpy as np
import pickle
import meshplot as mp
import open3d as o3d
import igl

df3d_folder = "../../../Downloads/data/deepfashion3dv2"
pose_folder = os.path.join(df3d_folder, "packed_pose")
pc_folder = os.path.join(df3d_folder, "point_cloud")
fl_folder = os.path.join(df3d_folder, "featureline_annotation")
mesh_folder = os.path.join(df3d_folder, "filtered_registered_mesh")

garment_id = 19 # the first in the short-sleeve upper category
garment_id = str(garment_id)

available_poses = [filename.split(".")[0] for filename in os.listdir(os.path.join(pose_folder, garment_id))]
available_poses = sorted(available_poses)
# available_poses

pose_id = available_poses[0]
fl_id = "_".join(pose_id.split("-"))
# pose_id, fl_id

# load pose and see if I can reconstruct the body
with open(os.path.join(pose_folder, garment_id, pose_id + ".pkl"), "rb") as f:
    packed_pose = pickle.load(f)
# packed_pose

model_n = smplx.create("./models/", model_type="smpl", gender="female",
                       global_orient=packed_pose["pose"][:3].reshape(1, 3),
                       body_pose=packed_pose["pose"][3:].reshape(1, 69),
                       transl=packed_pose["trans"].reshape(1, 3),
                       dtype=torch.float64)

output_n = model_n()
verts_n = output_n.vertices.detach().cpu().numpy().squeeze()
faces_n = model_n.faces

# newvn = (verts_n - offset) * packed_pose["scale"]# + offset
newvn = verts_n * packed_pose["scale"]

# next try to load the .ply files
pcd = o3d.io.read_point_cloud(os.path.join(pc_folder, garment_id, pose_id + ".ply"))

# visualize the garment mesh on body
v, f = igl.read_triangle_mesh(os.path.join(mesh_folder, pose_id, "model_cleaned.obj"))

p = mp.plot(verts_n, faces_n, c=np.array([0.8,1,0.8]))
p.add_mesh(v, f)
p.add_points(np.asarray(pcd.points))

If anyone could help me understand how to correctly load the data to view them that'll be great! Thanks!

Thank you for your work! The link is invalid (or inaccessible in China), please update your link again

Thanks for your great contribution.
After I downloaded the dataset, i could not find the multi-view real images as predicted on the Homepage.

Hi,

The google form has "Signed". May I know what this means?

Thanks,
Sai

          hi i would say there is still a problem i want to use them as a quantitative results to benchmark some models and the smpl is projecting outside the cloth and most of the clothes are actually for women but only the men smpl fits a in terms of height but the garment in colliding with it.

is there a better way such that the smpl doesnt cooincide?
if not the dataset wont be of any help to be

Originally posted by @msverma101 in #12 (comment)

Sorry, I cannot open the link of form. Can you give me another way to submit the form?

Hello, thanks for sharing the amazing dataset.

I have downloaded "deep_fashion_3d_point_cloud.rar" and successfully unziped the file with the passcode that provided via email.

However, I could not find the folder named "deep_fashion_3d_point_cloud_nnnotations" as mentioned in README page.

Maybe it hasn't been published yet?

Hi, just a quick question I know this was made with point cloud but Is there any way that I can use the result on Unity ?

I would like to know what software did you use to complete the annotation work?

Hi,
Thanks for making this excellent dataset public!

Since your SMPL's are not optimized for shape, we have garments that loosely/tightly fit the provided garment. If we want to optimize each and every SMPL model for shape parameters w.r.t garment, how can we do that?

Any tips/tricks will be helpful from anyone.

Hello,

According to your paper, there are four types of annotation in the dataset: "GT point cloud", "3D pose", "feature lines" and "multi-view real images".

I believe that first two annotations are already provided but couldn't find "feature lines" and "multi-view real images".

I'd very appreciate if you notify me of the expectation date for annotation data release.

Thanks

Hi @kv2000 , in the data provided, why are feature line and scan not aligned? How to solve this problem?

Generally speaking, it is only necessary to record the corresponding vertex index on the scanning data. Why is the absolute point cloud of feature lines provided here?

Hi,

Thanks for this amazing dataset. I was going through the dataset. Could you let me know about the following querry?

The smpl template is deformed based on featurelines vertices, using handle-based deformation.

Assuming, handle points are the boundary point on garment-template. Through segmentation masks of SMPL, we know which one is neckline, hemline etc.
However, how do we know the correspondence between such point on template to that of corresponding feature_line vertex locations, that are provided in the dataset. How to set hard-constraint if we don't know correspondence.