dmarcosg / dsac Goto Github PK

Python 100.00%

dsac's Introduction

Deep Structured Active Contours (DSAC)

This code allows to train a CNN model to predict a good map of penalizations for the different term of an Active Contour Model (ACM) such that the result gets close to a set of ground truth contours, as presented in [1] (to appear in CVPR 2018).

A preprint of the paper can be found in https://arxiv.org/pdf/1803.06329.pdf

Datasets

Vaihingen buildings

Bing Huts

Usage

Download and unzip the datasets. Modify the dataset paths in the main files and run them with Python 3. Requires Tensorflow 1.4.

Please contact me at [email protected] for questions and feedback.

[1] Marcos, D., Tuia, D., Kellenberger, B., Zhang, L., Bai, M., Liao, R. & Urtasun, R. (2018). Learning deep structured active contours end-to-end. arXiv preprint arXiv:1803.06329.

dsac's People

Contributors

Stargazers

Watchers

dsac's Issues

apply_gradients might not handle batches correctly

Talking about this line:
https://github.com/dmarcosg/DSAC/blob/master/main_vaihingen.py#L144

It seems to me that only the last image of the batch has an influence on the values of the gradients array. The values of these arrays are re-initialized for each j (i.e. each image of the batch) and only the last value is passed to apply_gradients. The first images of the batch hence seem to have no impact whatsoever on the weights of the network.

Therefore for batch_size of 1 everything works as expected but for other sizes, it won't.

ACM Initialization

Hi Diego,

Thank you for your great work! According to the paper, there are two methods to initialize the ACM. However, I was still wondering where exactly the ACM is initialized during the test phase.

active_contour_maps_GD_fast model file is not found

Hi,

Please help with active_contour_maps_GD_fast.py file. It's not available in branch.

Regards,
Raghav
[email protected]

difference in CNN models

Hi,

May I know what changes you have made in different CNN models in your code and how are they different from each other. Which one did you actually use in your implementation?
CNN, CNN_B, CNN_B_alpha, CNN_B_scalar.

Also, I want to know, is it possible to modify this code to include one more energy term?

Thanks.

TorontoCity implementation & csv generation

Hi Diego,

Thank you for your great work! I have downloaded ur codes and run it sucessfully but still have two questions. First, I did't find the implementation of TorontoCity even its dataset, could u pls upload it or provide the link? Second, how to generate the csv file just like in the bing and vaihingen dataset if I have some contour ground-truth.

Thanks alot and Best
Jie

I can only find l2 loss and Adam as CNN optimizer.

Please forgive if I have missed understanding the code, but it would be highly helpful if you could highlight these functions in your given code.

Thanks, hoping for your prompt reply.