talshaharabany / deepacm2d

Deep snake algorithm for 2D images based on ICLR2020

Deep Active Contours Model(ACM) for Buildings Segmentations

Deep snake algorithm for 2D images based on - [ICLR2020 paper(revisiting the results)](https://arxiv.org/abs/1912.00367) Architecture based on [Hardnet85](https://arxiv.org/abs/1909.00948) Data and weights = (https://drive.google.com/drive/folders/1fBSjPse3d8geV_iI3-PXV3x2qmLoUnzL?usp=sharing)

Get Started

To train a segmentation model :

python train_seg.py -bs 50 -WD 0.00005 -D_rate 3000 -task bing -opt sgd -lr 0.02 -nW 8

To train a ACM model :

python train.py -bs 25 -WD 0.00005 -D_rate 30 -it 2

To eval a segmentation model :

python eval.py -task viah -nP 100 -it 3 -a 0.4

To eval a ACM model :

 python eval_seg.py -task bing -nP 24 -it 2 -a 0.4

Results

Method	Viah mIoU	Bing mIoU
DARNet	88.24	75.29
DSAC	71.10	38.74
ours	90.33	75.53

deepacm2d's Issues

encounter problems about configuration of running env

the pkg is:
pytorch 1.7.0
torchvision0.8.0
cuda 11.4
ubuntu 18.04

As a green hand, I tried to install neural-renderer-pytorch in many ways but still failed. can anyone share the weights（bing/best/ACM.pt bing/best/SEG.pt ,or vaihingen's weights is ok）so i can check the "validation", thanks for any help!

btw @talshaharabany @evinpinar