search

junfu1115 / DANet

Dual Attention Network for Scene Segmentation (CVPR2019)
MIT License
2.41k stars 483 forks source link

DANet on pascal voc 2012 #59

Open PangYunsheng8 opened 5 years ago

PangYunsheng8 commented 5 years ago

Have you try to evaluate your DANet on pascal voc 2012? I did an experiment on that dataset but I did not get some gains.

gaowu9595 commented 5 years ago

@PangYunsheng8 Hi, can you tell me the results you have got?

PangYunsheng8 commented 5 years ago

@PangYunsheng8 Hi, can you tell me the results you have got?

MIOU is about 75%-76% on pascal voc 2012. I replaced PAM module to ASPP module, the miou is 79%.

Tramac commented 5 years ago

@gaowu9595 Hi, did you evaluate FCN-resnet50 on pascal voc 2012 (train:1464, val: 1449)? The mIoU is only 53%.

PangYunsheng8 commented 5 years ago

@gaowu9595 Hi, did you evaluate FCN-resnet50 on pascal voc 2012 (train:1464, val: 1449)? The mIoU is only 53%.

I evaluate FCN-resnet101 on pascal voc 2012, the miou is 75%. Note that the train sets include 10582 images, which is an augmented version on original pascal voc, and most of papers use this augmented voc dataset for training.

PangYunsheng8 commented 5 years ago

@PangYunsheng8 Hi, can you tell me the results you have got?

What's your results? DANet on pascal voc 2012.

Tramac commented 5 years ago

@PangYunsheng8 I have not tested DANet, I will test DANet-resnet50 on citys and post the result here soon. What's yours?

PangYunsheng8 commented 5 years ago

@PangYunsheng8 I have not tested DANet, I will test DANet-resnet50 on citys and post the result here soon. What's yours?

I only tested DANet on pascal voc 2012, because it costs too much times on citys, about 3-4days with two GPU.

Tramac commented 5 years ago

@PangYunsheng8 So, i just tested on pascal voc (train: 1464), it takes about 2~3 hours, the mIoU is low, i have no idea if the low mIoU caused by less data or my code.

Tramac commented 5 years ago

@PangYunsheng8 I downloaded the pascal aug dataset using ./scripts/prepare_pascal.py, but the train.txt just includes 8498 samples, how did you generate pascal aug dataset?

XUYUNYUN666 commented 5 years ago

@PangYunsheng8 Hi, can you tell me the results you have got?

MIOU is about 75%-76% on pascal voc 2012. I replaced PAM module to ASPP module, the miou is 79%.

wa....that is great! I just have one GPU xp, and so the batchsize is set to be 2 , the miou is only 0.61.....on cityscapes, what is your computer implement?

Exception-star commented 5 years ago

@PangYunsheng8 Hi, can you tell me the results you have got?

MIOU is about 75%-76% on pascal voc 2012. I replaced PAM module to ASPP module, the miou is 79%.

How do you train? what is your computer implement?