mightychaos / lkvolearner Goto Github PK
View Code? Open in Web Editor NEWLearning Depth from Monocular Videos using Direct Methods, CVPR 2018
License: BSD 3-Clause "New" or "Revised" License
Learning Depth from Monocular Videos using Direct Methods, CVPR 2018
License: BSD 3-Clause "New" or "Revised" License
Hi, @MightyChaos , great work!
Is there any reason to use exponential coordinates instead of Euler angles to represent camera rotation?
Does this alternative improve performance? Do you have some quantitative comparisons? Thank you!
I trained the model and tested each checkpoint I saved, and something seemed weird--- the test results were
abs_rel, sq_rel, rms, log_rms, a1, a2, a3
0.4429, 4.7569, 12.083, 0.588, 0.000, 0.303, 0.561
all the time.
What's wrong with it? Thanks.
Thank you very much for providing this interesting work.
I have a question about the scale of the output depth. As declared in the paper, when we multiple a scale on to the depth output (as well as the pose) the loss(L_{ap}) won't change. So how can we make sure that the 1/(output of vgg_depth_net) have the same scale as the ground truth when doing the evaluation.
Very interesting work!
I have one question about our work. There is no report of the pose evaluation result in your paper and code. Do you evaluate the estimated pose with ground truth?
It seems missing a rot_angle_cos() in the row 121 in the Twist2Mat module
Thanks for your share of codes.
Could you please show the codes of testing on Make3D? I try to reproduce the results, but I can't get the good one. Now my results are
abs_rel | sq_rel | rmse | rmse_log |
0.446 , 6.667 , 9.839 , 0.212
The results are too bad comparing to the paper.
Could you please provide the codes? Thanks very much.
Thanks for the nice work. I was reading the code to better understand the paper. But I don't understand about this line:
LKVOLearner/src/DirectVOLayer.py
Line 330 in 4947803
If I understand it correctly, it should be the Jacobian of x with respect to p, where x is the pixel location in the image space, and p is the lie vector for the transform in se3. I tried to derive the Jacobian to get the same form as your code, but I have not figured it out yet. Is this part illustrated anywhere in supplementary? Can you share your derivation for this formula, or refer me to any document?
Hi,
In your paper you point out that batch size is set to 1 because of implementation, I guess it points to the implementation of DDVO algorithm. But I found that when training posenet, batch size is constrained to 1 in SfMKernel. Is that necessary?
assert(frames.size(0) == 1 and frames.dim() == 5)
And when I train DDVO without posenet it took me 3.5 hours to train half an epoch. I wonder is that a long time normal?
Thanks
Hello,
Thank you for releasing your code and your model.
I see that you provide instructions for both training only DDVO and the version where you initialize DDVO with PoseNet. However I can't find where to download the weights to initialize the finetuning. In your code, it corresponds to what is called depth_net.pth
and pose_net.pth
. Is there a link from where I can download them please ?
Thank you in advance.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.