Open Xianbo-X opened 2 years ago
HANet: https://github.com/shachoi/HANet R2U-Net + HANet: https://github.com/tomasamado/cityscapes-image-segmentation DSNet: https://arxiv.org/abs/1904.05022
add 2 recurrent layer (epoch=10, batch=4) ---> worse result
R2Unet64 (code from github, epoch=15, batch=4)---> worse result reference: https://github.com/tomasamado/cityscapes-image-segmentation/blob/main/Vision_task_2.ipynb
cat_unet (epoch=10, batch=4)
cat_unet, 1/2 in_channels in Up() (epoch=10, batch=4) ---> better and faster than (3), but slower than baseline (7min)
@Xianbo-X
OK. I got a similar result for the Unet+Positional mask. I have a new idea. We can cut the whole image into pieces to obtain better results. Ie. Cut the whole images into pieces and train the network on pieces.
First version of RNN- 1 epoch
Only recognize the abstract shape of roads
On trainset:
Unet-based RNN 图像中可以看由于切割出现的变形。 我感觉如果调整一下结构,延长一下epoch或许有效果
和baseline比较可以看出切割训练在某些细节处处理更好。 数字为IoU.可以看出,在某些例子上RNN表现更好 在绿色像素尤其表现良好
人多的时候尤其差
我做了一个简单的unet-based的RNN。没有仔细调整。然后放了些结果在github上。你看看调整下。我觉着这个方向是可以的。
调整了model,现在可支持gpu加速,在notebook中需要更改以下部分:
Why should we do cuda_tri=False for cpu model?
Baseline: UNet Improvement:
DSnet concatenate method @LYYYYx
Upsampling indices (From SegNet). https://github.com/yassouali/pytorch-segmentation/blob/master/models/segnet.py @Xianbo-X
RNN @LYYYYx
Transformer (--> HANet) @Xianbo-X
CFR*
Further part
Attention: HANet