Comments (7)
In our paper loss of alpha, F, and the last term is L1 loss (absolute distances). In the DIM paper it is L2 loss. DIM shows how to make the gradient differentiable. This is not needed anymore as pytorch can handle it through auto-diff. Loss in eq5 of LFM is same as the 2nd term in our paper.
from background-matting.
Empirically I saw that to be effective. One explanation, learning matte is easier than the foreground, as alpha matte is mostly 0 or 1 and continuous at only certain regions. Often these weights are chosen empirically to make sure all loss components are converging at the same rate.
from background-matting.
Can you tell me what you mean by inspiration? The supervised loss was very similar to previous works, nothing new. Gradient loss was used by Late Fusion matting, it helps to get sharper edges in the alpha, else the network is biased towards smoother edges
from background-matting.
I looked into the Late Fusion matting paper but I could not find the loss that you used so I thought your loss was an original one and wanted to know how you made it.
from background-matting.
The supervised L1 loss is similar to Deep Image Matting paper (they have losses over alpha and composition). We did not attempt to make it differentiable and use simple L1 loss, also add a loss over the foreground F. Note that many existing matting works do not predict the foreground layer F and simply use the image I while compositing. This causes artifacts (see our paper appendix).
The gradient loss is added from Late Fusion Matting paper (eq 5)
from background-matting.
So you are saying this loss (from your paper):
is similar to these losses (from the DIM paper):
and
And the term with the gradient is from the Late Fusion Matting paper:
Thank you for your time and patience. Just want to be sure. I'm admittedly not good at this math. Its just that these equation don't seem to match...at least on visual inspection.
from background-matting.
Thank you so much! This has brought alot of clarity! One last question... why does the loss of F get more weight/coefficient of 2.
from background-matting.
Related Issues (20)
- Can I skip the Alignment step? and How? HOT 1
- issue in JPG inputs HOT 1
- loss log HOT 1
- About the training time HOT 1
- matting performance problem HOT 1
- Real time results with constant background HOT 1
- SAD in the paper
- Evaluation on Synthetic-Composite Adobe Dataset
- [error] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR HOT 1
- Background Image HOT 3
- How to test after training on unlabeled real videos
- Camera Model For Fix-Camera dataset
- 这个对抗损失的意义是什么?怎么理解? HOT 1
- The download link for pre-trained model is failed! HOT 1
- Possible Memory Leak with AdobeDataAffineHR
- 是否可以训练他来分割别的东西,比如说超市商品?
- Could he be trained to split other things, such as supermarket goods?
- Using sample data does not achieve the effect in the paper
- Colab notebook is no longer works HOT 1
- My colab does not have cuda 10.0. There is only 12.0 available now. Can we still run this model?
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 background-matting.