zhimingzo/modulated-gcn

Modulated Graph Convolutional Network for 3D Human Pose Estimation

modulated-gcn's Issues

Performance

Thanks for your novelty work, I retrain the network, but I can only get 51.33 on average MPJPE, which is lower than your paper's performance of 49.4. You use the refinement module, if it is not fair that other methods don't use the refinement module?

Network architecture question.

Hi, thanks to your great work and sharing the code.

I have a question about your network architecture.
In your paper, you said "All the graph convolutional layers are followed by batch normalization and a ReLU activation function except for the last one".
However, in your code ./models/modulated_gcn.py (GT) which likes as below:

 def forward(self, x):

        x = self.gconv(x).transpose(1, 2)
        x = self.bn(x).transpose(1, 2)
        if self.dropout is not None:
            x = self.dropout(self.relu(x))

        x = self.relu(x)
        return x 

the architecture in your code looks like {gconv-bn-relu-dropout-relu}, on the other hand, in your paper
{gconv-bn-relu-dropout}.

If you missing else: operation in your code?
Please check which one is right.

Thank you in advance.

About Reproduction of the numbers in the paper!

This is my training result: 52.73.
The result given in Table 5 in your paper is 49.4.
I don't know how to solve this problem, I hope you can help me solve this problem.

Data processing and code structure

Hello! Thank you very much for your work, which has helped me a lot! I want to ask a few questions about the preparation of the dataset and the code structure.

First of all, where is the data set directory? It is related to Modulated-GCN_ benchmark、Modulated_ GCN_ Gt Peer directory code? Because CPN 2D detection is placed in this directory. Is the human 3.6 dataset also placed under the dataset?

Second, Modulated-GCN_ Benchmark and Modulated-GCN_ What are the two specific differences between gt?

Third, Modulated-GCN_ The Google netdisk where the fine-tuning posture is located in the readme of data under gt cannot be accessed.

Looking forward to your reply, thank you again! @ZhimingZo

python main_ Gcn.py execution error

Hello, thank you for your position. I'm running python main_ The size of Tensor a (16) must match the size of Tensor b (17) at non-singleton dimension 1 appears when the gcn.py command is executed? Looking forward to your reply, thank you. @ZhimingZo

About l1 Norm and l2 Norm Implementation

Modulated-GCN/Modulated_GCN/Modulated-GCN_benchmark/train_graph_time.py

Line 122 in fd15af1

loss_gt = (1-opt.norm) *eval_cal.mpjpe(pred_out, out_target) + opt.norm*criterion_L1(pred_out, out_target)

Hi, could u please say where you implemented the l2 norm...in the above line I can only see the l1 norm is used.

And if I'm not wrong it won't calculate the below line because in ur case pad is always 0.

Modulated-GCN/Modulated_GCN/Modulated-GCN_benchmark/train_graph_time.py

Line 135 in fd15af1

loss_diff = criterion_mse(diff, Variable(torch.zeros(diff.size()), requires_grad=False).cuda())

could you please explain where you are calculating the l2 norm?

Testing on MPI-INF-3DHP

Thanks for the great work. Could you please tell me how can I test on the MPI-INF-3DHP dataset & reproduce the result of Table 7?

Weight Modulation

hello, thanks for your great work! Here are some my confusions. What's the difference between Weight Modulation and sem_ch_graph_conv in SemGCN? In table 2, what's the difference between sparsity and l1-loss? In your code implementation, the sparsity constraint is the same as what mentioned in Sec3.4. Looking forward your reply,thanks!

Purpose of E and 1-E

Could you please say why you are taking E in

Modulated-GCN/Modulated_GCN/Modulated-GCN_benchmark/models/modulated_gcn_conv.py

Line 44 in fd15af1

E = torch.eye(adj.size(0), dtype=torch.float).to(input.device)

and what is the difference between doing E and 1-E in the below line

Modulated-GCN/Modulated_GCN/Modulated-GCN_benchmark/models/modulated_gcn_conv.py

Line 46 in fd15af1

output = torch.matmul(adj * E, self.M*h0) + torch.matmul(adj * (1 - E), self.M*h1)

Time series data

Have you tried the Modulated GCN model for time series data ?

h36m_skeleton_joints_group

Hi! Thanks for your work.

I am trying to use your work on other datasets and am confused about the definition of 'h36m_skeleton_joints_group'.

h36m_skeleton_joints_group = [[2, 3], [5, 6], [1, 4], [0, 7], [8, 9], [14, 15], [11, 12], [10, [13]]

For example, 5 is r-top and 6 is l-hip... Why do you group these two joints? How can I define the 'h36m_skeleton_joints_group' on other dataset? Thank you!!