search

juhongm999 / hsnet

Official PyTorch Implementation of Hypercorrelation Squeeze for Few-Shot Segmentation, ICCV 2021
231 stars 43 forks source link

牛逼啊,直接干这么高让我们无路可走 #1

Closed ily666666 closed 3 years ago

ily666666 commented 3 years ago

牛逼啊,直接干这么高让我们无路可走

zbf1991 commented 3 years ago

Seems that they use 400*400 size during influence rather than original image size

ily666666 commented 3 years ago

Seems that they use 400*400 size during influence rather than original image size

所以,你认为如果把尺寸改为224效果会降低吗

zbf1991 commented 3 years ago

see the performance of PFENet on COCO, a fixed size generates a higher performance than the original size.

ily666666 commented 3 years ago

对于FSS数据集,他们本来就有一个固定的尺寸224224,变换到400400 影响应该不大. 另外,我刚刚又看了一下PFENet的论文,这篇论文(HSNet)表格中PFENet的数据也是固定尺寸473*473得到的. 不知道别的什么情况,但是有一定的提升这一点是毋庸置疑的.

juhongm999 commented 3 years ago

@ily666666 @zbf1991

During both training and evaluation, we use fixed image size of 400 x 400 following recent state of the arts. For example, PPNet and RePRI both use spatial size of 417 x 417 for the input images. PFENet uses both fixed (473 x 473) and original image sizes in its evaluation. In our experiments, we found that larger image sizes typically (> 417) result in better mIoU results but we use image size of 400 x 400 as the image sizes of 417x417 and 473x473 seemed random numbers to us and our model with the size of 400 x 400 already surpassed previous state of the arts. To see the primary source of performance improvements, please refer ablation study in Section 5.2 of our paper.

tymatfd commented 2 years ago

牛逼啊,直接干这么高让

牛逼啊,直接干这么高让我们无路可走

握手，看到这结果当时心都凉了。现在还好已经有至少两个工作比他还要高了，我可以彻底躺平了