Performance is lower than the results you report.

[zhouyuan888888]

By using the default config you provide, I have achieved 79.206 % (your result is 79.916 %) based on deeplabv3+res101 on the pascalvoc validation set. Can you tell me why this happened? By the way:

1. I have used the pretrained res-101.
2. My environment is python3.6, cuda8.0, pytorch0.4.1,
3. The augment pascal voc dataset is downloaded from DrSleep
4. I re-train the model without any modifications for the code you provide.

How can I do to achieve the performance you report in the README.? Thank you for your help~

[YudeWang]

@zhouyuan888888
Have you use deeplabv3+vocfinetuning in experiment folder to refine your 79.2.6% model? The annotation of VOC train set is better than trainaug set. Experiment of deeplabv3+vocfinetuning will finetune on VOC train set with a small learning rate and further improve mIoU.

[zhouyuan888888]

Thank you for your reply : ). I know finetuning the model on the original pascal voc train set can improve accuracy. But in README you pointed out that the performance of 79.155% and 79.916 % is achieved without finetuning. So I have tried many times to reproduce your results but failed. T_T, So, if I want to reproduce your results, what can I do? It is so strange because I have not modified any your codes or configurations.

Thank you so much for your answering! :)

[YudeWang]

@zhouyuan888888
deeplabv3+res101 achieve 79.155% and deeplabv3+xception achieve 79.945%
I find you have said that pretrained res101 model was used. The comparison is correct? your-79.206%>my-79.155%

[zhouyuan888888]

Yes, I used the pretrained res101 model. So I compare our results based on the res101 version. 79.206 % and 79.916 % are all evaluated by using multi-scale and flip strategies. :)

[YudeWang]

@zhouyuan888888
Maybe it is caused by random seed? The code I released does not fix it. And what about single scale test?

[zhouyuan888888]

I think the gap about 0.7% is a little large. For single-scale test without flipping, I only got about 78.014 % MIoU (yours is about 79.155% ) T_T. By the way, have you ever used any other tracks to further import accuracy?

You are so kind, thank you for your quick reply!!!

[YudeWang]

@zhouyuan888888
There is no other tricks. Please finetuning on train set (the paper includes that), or select better super-parameters by yourself. Dense CRF also can be considered.

[JingyuLi-code]

By using the default config you provide, I have achieved 79.206 % (your result is 79.916 %) based on deeplabv3+res101 on the pascalvoc validation set. Can you tell me why this happened? By the way:

1. I have used the pretrained res-101.
2. My environment is python3.6, cuda8.0, pytorch0.4.1,
3. The augment pascal voc dataset is downloaded from DrSleep
4. I re-train the model without any modifications for the code you provide.

How can I do to achieve the performance you report in the README.? Thank you for your help~

Hello, 我和你上述保持同样的参数，包括pretrained res-101.augment pascal voc dataset is downloaded from DrSleep，以及 多尺度测试和裁剪,在4块RTX2080Ti实验。
在训练46个epoch]后测试，得到以下结果
backbound: 94.368% aeroplane: 92.200%
bicycle: 43.120% bird: 89.926%
boat: 71.572% bottle: 82.099%
bus: 95.505% car: 89.099%
cat: 92.570% chair: 40.066%
cow: 89.992% diningtable: 56.415%
dog: 87.355% horse: 86.244%
motorbike: 87.790% person: 86.892%
pottedplant: 62.389% sheep: 88.140%
sofa: 47.990% train: 87.240%
tvmonitor: 78.863%

   mIoU: 78.564%

在46epoch 我没得到满意的结果。所以我尝试增大epoch,保持github其他参数不变，将epoch改为60，训练60个epoch后测试：
backbound: 94.560% aeroplane: 93.138%
bicycle: 42.715% bird: 89.845%
boat: 76.035% bottle: 80.670%
bus: 95.354% car: 89.574%
cat: 94.906% chair: 41.242%
cow: 87.774% diningtable: 59.855%
dog: 89.180% horse: 86.608%
motorbike: 86.478% person: 87.563%
pottedplant: 61.951% sheep: 89.305%
sofa: 47.412% train: 89.707%
tvmonitor: 77.503%

   mIoU: 79.113%

在itr35000测试：
backbound: 94.661% aeroplane: 93.178%
bicycle: 42.561% bird: 89.823%
boat: 76.099% bottle: 80.941%
bus: 95.403% car: 89.597%
cat: 94.806% chair: 40.680%
cow: 87.209% diningtable: 60.366%
dog: 88.884% horse: 86.239%
motorbike: 86.410% person: 87.537%
pottedplant: 62.798% sheep: 89.348%
sofa: 49.260% train: 90.454%
tvmonitor: 77.807%

   mIoU: 79.241%

然而，在vocfinetune 训练60epoch结果会很差：
finetune
backbound: 93.994% aeroplane: 91.886%
bicycle: 43.208% bird: 89.967%
boat: 71.949% bottle: 78.653%
bus: 94.821% car: 88.435%
cat: 94.454% chair: 38.411%
cow: 88.571% diningtable: 45.917%
dog: 88.607% horse: 86.051%
motorbike: 85.921% person: 87.293%
pottedplant: 60.941% sheep: 87.682%
sofa: 44.132% train: 85.685%
tvmonitor: 76.548%

   mIoU: 77.292%

所以，请问你后续有没有进行超参数的调试以达到工程所达到的结果？thanks!

[CharlesPikachu]

maybe you can try this: https://github.com/SegmentationBLWX/sssegmentation