search

openseg-group / OCNet.pytorch

Please choose the openseg.pytorch project for the updated code that achieve SOTA on 6 benchmarks!
MIT License
812 stars 128 forks source link

why not try VOC dataset #29

Closed idealwei closed 5 years ago

idealwei commented 5 years ago

I've read your paper and I know your task is scene parsing. But why not experimenting on Pascal VOC datasets ? Will OCNet achive SoTA on VOC too ?

PkuRainBow commented 5 years ago

We don't have a good baseline on VOC dataset.

I guess OCNet generalizes well to other dataset.

idealwei commented 5 years ago

Can you try to explain why ? Did you train VOC like deeplabv3plus does ? It's very hard to finetune the model to get such high scores on VOC dataset. And i want to know if OCNet performs better than PSPNet or Deeplabv3 on VOC datasets. I remember the author of Deeplabv3plus remove average global pooling of ASPP and get better results on cityscapes datasets and prove image pooing is more suitable to VOC dataset. I think if you can also get a better score on pascal voc dataset you can get an oral.

PkuRainBow commented 5 years ago

@idealwei Thanks for your advice.

In fact, we also want to validate our method on VOC. But our baseline is far from the current state-of-the-art method.

It requires much more tricks to achieve good performance on VOC.

It would be great if you have access to Pytorch based baseline on VOC dataset. We can conduct the related experiments quickly.

We believe that Object Context is a very general concept towards better segmentation performance.

idealwei commented 5 years ago

@PkuRainBow. You can refer to GluonCV and here is its PyTorch implementation. Hope that can help.