Comments (5)
Thank you.
Question 2. I got definition in Social-LSTM[1], a joint of several independent 2-dimension Gaussian distributions.
Question 1. In ST-GCN[2] model, kernel_size
is for spatial convolution on graph, where adjacency matrix is time-invariant. If my understanding is not wrong, it is a learnable kernel just like in regular CNNs. But in your paper, adjacency matrix is time-variant and non-learnable. So I think torch.einsum('nctv,tvw->nctw', (x, A))
is better. And parameter kernel_size
can be removed.
The complete code is as follows. I'm testing whether this change will influence the result.
class ConvTemporalGraphical(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
t_kernel_size=1,
t_stride=1,
t_padding=0,
t_dilation=1,
bias=True):
super(ConvTemporalGraphical,self).__init__()
self.kernel_size = kernel_size
self.conv = nn.Conv2d(
in_channels,
out_channels,
kernel_size=(t_kernel_size, 1),
padding=(t_padding, 0),
stride=(t_stride, 1),
dilation=(t_dilation, 1),
bias=bias)
def forward(self, x, A):
assert A.size(0) == self.kernel_size
x = self.conv(x)
x = torch.einsum('nctv,tvw->nctw', (x, A))
return x.contiguous(), A
[1] Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., & Savarese, S. (2016). Social LSTM: Human trajectory prediction in crowded spaces. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2016-December, 961–971.
[2] Yan, S., Xiong, Y., & Lin, D. (2018). Spatial temporal graph convolutional networks for skeleton-based action recognition. 32nd AAAI Conference on Artificial Intelligence, AAAI 2018, 7444–7452.
from social-stgcnn.
Hi,
- We are not using regular TCN in our work, we just refer to the concept and correlate it with TXPCNN layer that treats the temporal dimension as a feature channel; unlike TCN that treats temporal data as pixel values.
Theeinsum
is Einstein sum which is a concept you can google for it; What we are trying to do here is to collapse the graph sequences into a single representation of <time,ped,features> using the graphs and their corresponding adjacency matrix (A); in other terms we weight the features from neighbor pedestrians to a specific pedestrian using A.
2, Social-STGCNN is not a deterministic model, if you refer to the loss function in the paper we model the trajectory as a bi-variate gaussian distribution and predict the 5 parameters of each trajectory in time which are mean_x, mean_y,variance_x,variance_y and correlation_xy.
By predicting the distribution, you can sample multiple trajectories, in our testing we sample 20 trajectories as this was a community standard for these kind of problems.
Thanks
from social-stgcnn.
Hi,
Thanks for your notice on this; I re-ran the experiments again and obtained similar results as per your suggestions and it makes sense. I also updated the repo accordingly.
from social-stgcnn.
Hi,
I wonder which commit is the vesion of your paper published in CVPR 2020.
Thanks!
from social-stgcnn.
@d-zh https://github.com/abduallahmohamed/Social-STGCNN/tree/ebd57aaf34d84763825d05cf9d4eff738d8c96bb
from social-stgcnn.
Related Issues (20)
- anorm problem about step_rel and step_ HOT 4
- Data visualization HOT 3
- Why does the loss value change from positive to negative HOT 1
- PROBLEM About TCN HOT 1
- could min-pedestrian greater than 1? HOT 1
- How to handle trajectories with missing frames? HOT 2
- Visualize trajectory in picture HOT 6
- Details about data processing HOT 3
- Questions about data process
- I am very curious about the element in your proposed adjacency matrix
- I am very curious about the element in your proposed adjacency matrix HOT 3
- When I test my own model, some datasets work fine, but the univ dataset reports errors. The specific error message is as follows. HOT 1
- details about dataset .txt file encoding type. HOT 1
- A question of determining linear or nonlinear
- During visualization, the jupyter notebook encountered the following error
- Details about Gaussian path visualization
- Question about the code in basic gcn unit
- final predicted trajectory data
- This work is very cool. As a beginner in learning trajectory intention prediction, I would like to ask some questions.
- make our own dataset
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from social-stgcnn.