Hi, could you give a demo of using test_color.py and train_color.py? Is there any special requirement on the data?

I made several attempts on training and testing the color predictor, I can see that the data in the .obj file is in a nx6 format. Is each row composed of vertex coordinates and the corresponding RGB value(in (0,1))?

Could you recommend a proper way to render the mesh with color using the .obj file?

Many thanks.

I wanted to know if there is a reason why the generated animated mesh is flattened/smooth with only the low frequency details even though the input mesh has high frequency features. Is there a parameter or option to get the results in high frequency fidelity?

Input Mesh - Generated Mesh -

Dear authors, thanks for your amazing work. I have a questions about minimal bofy: I found running test_ scanimate.py requires a canonical pose of 6890 vertices, but the number of vertices of cano.obj obtained by running the training code is greater than 6890, so it cannot be run on the test code. How to get minimal body of my own data for test_scanimate.py?

for training you mention that we need 3 things - input scan, SMPL parameters of the scan and a T-pose SMPL mesh. This works for SMPL. You have mentioned extensions to testing with smplx, but do you know what are the things I need to change to train with the SMPLx model? Do I need to change default.yaml and example.yaml?

Dear authors,

Many thanks for sharing such an amazing work. Without training, would it be possible to test this tool on 3D scans of unseen subjects in the training? The training data consisting of pairs of (raw scans, motion parameters) is not straightforward to obtain for a given subject.

Thanks!

Dear authors, thanks for your amazing work. I am trying to follow your method, and have some questions about \data usage:

1. Which subjects in CAPE did you use for the reported metrics? Are they beyond the publically available 20 raw scan sequences of 4 subjects in CAPE?

2. You mentioned in the paper that "the motion sequences are randomly split into training (80%) and test (20%) sets". Do you mean "80% frames" or "80% sequences"? Is it true that at least a part of each sequence is used for training?

Thank you!

code: model = smpl.create(opt['data']['smpl_dir'], model_type='smpl_vitruvian', gender=opt['data']['smpl_gender'], use_face_contour=False, ext='npz').to(cuda)
The above line of code reports an error.

File "/root/anaconda3/envs/pytorch3d/lib/python3.7/site-packages/smpl-0.1.13-py3.7.egg/smpl/body_models.py", line 94, in create ValueError: Unknown model type smpl_vitruvian, exiting!

Is this a smpl installation error?

After training with the data of example_03375_shortlong, I obtain unsatisfactory results. There are many noise points. Further, the left upper leg is weird. Is it because the number of frames (16 only) is not enough? Do you have suggestions to obtain better results?

想请问一下，如果设置多GPU 并行处理？ 在代码中设置： os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ['CUDA_VISIBLE_DEVICES'] = "4,6"; 开头： cuda = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 中间： if torch.cuda.device_count() > 1: print("Let's use", torch.cuda.device_count(), "GPUs!") model = nn.DataParallel(model); model = model.module model.to(cuda); else: model.to(cuda);

但是不起作用！

CapeDataset.py/load_trimesh()

This function uses a fixed value of ground level(line 60). This creates problem for new datasets if the ground level for the new dataset is much lower.

1. How to fit a SMPL model to the input 3D scan/mesh - how to extract the pose and shape parameters of the 3D scans is this done along-side scan generation or can it be done posterior to any arbitrary 3D mesh/3d scan
2. Do you intend on providing any tools/suggest methods to extract the SMPL minimal and pose parameters for a given 3D mesh

Hi, @shunsukesaito. I am trying to follow your amazing work. However, I encountered some problems with the evaluation.

Would you mind releasing the Evaluation and Metric Code of Table 1 (Ds2m (cm) and Dn of Int. and Ex.)?

Hi, I'm training SCANimate on ClothSeq dataset provided by Neural-GIF(https://github.com/garvita-tiwari/neuralgif). However, I met some problems and got bad results.

Here are the logs of the first stage when training on a Cape sequence:

And here are the logs of the first stage when training on a ClothSeq sequence:

I notice that the pretraining of the skinning net converges very quickly on the Cape sequence, while the loss drops slowly on the ClothSeq sequence. I have checked the minimal_body file, the scans and SMPL parameters of each frame, all of them have the same format as the Cape dataset and are aligned well. So I'm wondering what the problem is? Do I need just more epochs in the first stage, or something else? I hope you can give me some advice, thank you!