Xianbo-X / 5LSM0_Cityscapes_challenge

Project repository for final assignment of 5LSM0
0 stars 0 forks source link

Roadmap Modeling selection and implementation #7

Open Xianbo-X opened 2 years ago

Xianbo-X commented 2 years ago

Baseline: UNet Improvement:

LYYYYx commented 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

LYYYYx commented 2 years ago
  1. add 2 recurrent layer (epoch=10, batch=4) ---> worse result rnn2_epoch10_batch4

  2. R2Unet64 (code from github, epoch=15, batch=4)---> worse result reference: https://github.com/tomasamado/cityscapes-image-segmentation/blob/main/Vision_task_2.ipynb r2unet64_epoch15_batch4

  3. cat_unet (epoch=10, batch=4) cat_epoch10_batch4

  4. cat_unet, 1/2 in_channels in Up() (epoch=10, batch=4) ---> better and faster than (3), but slower than baseline (7min) cat_in2_epoch10_batch4

@Xianbo-X

Xianbo-X commented 2 years ago

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.

Xianbo-X commented 2 years ago

First version of RNN- 1 epoch

Only recognize the abstract shape of roads image image

On trainset: image

Xianbo-X commented 2 years ago

Unet-based RNN image 图像中可以看由于切割出现的变形。 image image image image 我感觉如果调整一下结构,延长一下epoch或许有效果

Xianbo-X commented 2 years ago

和baseline比较可以看出切割训练在某些细节处处理更好。 image 数字为IoU.可以看出,在某些例子上RNN表现更好 image image 在绿色像素尤其表现良好

人多的时候尤其差

Xianbo-X commented 2 years ago

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。

LYYYYx commented 2 years ago

调整了model,现在可支持gpu加速,在notebook中需要更改以下部分:

  1. Trainer -> train_epoch()

image

  1. Trainer -> val_epoch()

image

  1. testing_training()

image

Xianbo-X commented 2 years ago

Why should we do cuda_tri=False for cpu model?