Comments (8)
You are right.
By math, they will simply be cancelled due to the logarithms.
That ugly trick is just our empirical choice and will be improved in next release.
In fact, we are preparing new release due to inconsistency (small difference in performance) between currently relased model and the paper.
Thanks for kindly pointing this. =)
from bts.
Thanks for the code. I like the idea:) :)
Now that I'm looking at the code there is no beta implemented in the code. I was wondering if you removed beta in this provided code. I'm trying to reproduce your result that you submitted to the benchmark. Please let me know if something is missing for that.
I was also wondering if the model provided for eigen is the same model as the one you submitted to KITTI benchmark? If not can you mention the changes you did for the model you submitted to the benchmark ?
If you don't mind here is some inconsistencies I noticed if you wanna edit your paper:
In default parameters for eigen split(i.e. batch size and rotate degree)
and also I think n3 should be passed to a separate sigmoid (not tanh) and then n1,n2 (after tanh) and n3 (after sigmoid) be passed to L2 normalization. Nice trick to avoid minus of the normals btw :) You just need to update the figure I think.
Thanks,
Ali
from bts.
Hello, nice to hearing from you again. =)
As I mentioned before, I updated the code not to include \beta in the training loss, and updated the model file as well trained with the modified loss function.
As a matter of fact, the model submitted to the KITTI online benchmark is different from trained with Eigen split. The benchmark version was trained with sampled examples (about 57000 rgb and gt pairs) from whole dataset, but parameters are almost same: batch_size 16, num_gpus 4, num_epochs 50, input_height 352, input_width 704, do_kb_crop, do_random_rotate, degree 1.0.
The model for the Eigen split is trained with the parameters in "arguments_train_eigen.txt", only batch_size and num_gpus are different (actually they were 16 and 4, respectively) because I assume most of users will use our code with only one gpu.
Finally, I have to say that this work is still in progress.
I mean, before publication (not a preprint in arXiv) our code and paper will be updated.
This is why we made our code and paper publicly available, this kind of very constructive review from you! =)
Thanks,
Jin Han
from bts.
Closing this issue due to inactivity
from bts.
Hi, is the model submitted to the KITTI online benchmark trained on the official " annotated depth maps data set (14 GB)", which only contains 43000 rgb images.
from bts.
Hi, is the model submitted to the KITTI online benchmark trained on the official " annotated depth maps data set (14 GB)", which only contains 43000 rgb images.
from bts.
@luohongcheng Yes. The submitted model was finetued using only the official ground truth depth maps with the backbone network pretrained using ImageNet.
from bts.
Would you please share your file lists selected from the official training set? Thanks.
from bts.
Related Issues (20)
- MobileNetV2 HOT 2
- Question about the image size during training and evaluation HOT 1
- Asking for some advice on KITTI 3D det usage. HOT 2
- ValueError: batch_size should be a positive integer value, but got batch_size=0 HOT 3
- Predicting random images other than KITTI & NYU, when the model is trained on either of the dataset
- Consultation and advice for bts License
- Testing on single image from KITTI dataset
- Something wrong with kitti gt dataset HOT 1
- Inconsistent results on Online evaluation and Test (might caused by benchmark crop operation) HOT 3
- Dear author,Have you ever tried to use bts for multiscale output?
- Question about Ray-Plane Intersection in LPG
- Questions on the code organization
- Pretrain models
- How were the files selected in nyudepthv2_train_files_with_gt.txt? HOT 1
- Why does the depth output has smears on the left and right sides?
- what is the focal HOT 2
- Single machine with multiple gpus,but I only want one...
- Error in filenames file while training
- Can you please tell me how to figure out the actual depth value(in metres or centimeteres) at any (x,y) coordinate of the image that is being predicted by the model? HOT 2
- nyu test dataset HOT 1
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 bts.