search

leeyeehoo / CSRNet-pytorch

CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes
642 stars 259 forks source link

Perform data augmentation in training and normalize correctly at test… #9

Open covix opened 5 years ago

covix commented 5 years ago

Working with the code I noticed that data augmentation (i.e. random crop and random flip) was not performed, and that in val.ipynb data normalisation at test time was not working correctly. With these changes I was able to perform MAE: 11.12 on ShanghaiTech Part B, not far from what reported in the paper.

leeyeehoo commented 5 years ago

I'll check it today. Sorry for the delay.

vlad3996 commented 5 years ago

I've obtained 10.42 MAE and 16.89 MSE on part B without augmentation but it takes about 20 hours on GTX 1080Ti. So maybe augmentation is not needed at all.

covix commented 5 years ago

Probably part B is a big enough dataset, in which data augmentation does not make much difference. I tried CSRNet with a smaller dataset, and data augmentation do makes a lot of difference, it improves MAE from ~20 to ~10.

eain3314 commented 5 years ago

@vlad3996 请问MSE怎么加上的？谢谢