Couldn't reproduce the semantic segmentation experiment about vitae-transformer-remote-sensing HOT 11 CLOSED

The demo/*.ipynb are carried in the original mmsegmentation, and we haven't used these yet to test our models.

So maybe your difficuty is how to prepare the potsdam or isaid datasets? For potsdam dataset, you need to download it from: https://www.isprs.org/education/benchmarks/UrbanSemLab/default.aspx, and clip image to patches manually.
For isaid dataset, you need to download them from: https://captain-whu.github.io/iSAID/dataset.html. Then, you also need to clip them. When preparing the isaid dataset, the officials has provided the Development Kit. For using details you can refer to these repos: https://github.com/Z-Zheng/FarSeg; https://github.com/lxtGH/PFSegNets

Good luck!

from vitae-transformer-remote-sensing.

In this repo, we modify the original mmsegmentation and we have registered the ViTAE in https://github.com/ViTAE-Transformer/ViTAE-Transformer-Remote-Sensing/blob/main/Semantic%20Segmentation/mmseg/models/backbones/__init__.py

from vitae-transformer-remote-sensing.

Ha? The relevant remote sensing pretrained models, including RSP-ResNet-50-E300, RSP-Swin-T-E300 and RSP-ViTAEv2-S-E100 have already been released in the home page of this repo, please see readme, that is why we created this project.

from vitae-transformer-remote-sensing.

No, the MillionAID is a large-scale dataset for remote sensing scene classification, it doesn't contain any pixel-level labels. Our purpose is to transfer the weights trained from classification task to a series of remote sensing downstream tasks, such as aerial segmentation and detection.

from vitae-transformer-remote-sensing.

Oh, we have uploaded the trained upernet weights (fine tuning on potsdam and isaid) in the readme.md of segmentation part, maybe you can manully extract the corresponding weights of backbone part.

from vitae-transformer-remote-sensing.

I'm stuck at this point:
KeyError: "EncoderDecoder: 'ViTAE_Window_NoShift_basic is not in the models registry'"
when executing init_segmentor()
I didn't face it on the another models, only on ViTae.

Could you also provide this pre-training of the model? The one from now-MillionAID

from vitae-transformer-remote-sensing.

Which file is this error report from？

from vitae-transformer-remote-sensing.

In this repo, we modify the original mmsegmentation and we have registered the ViTAE in https://github.com/ViTAE-Transformer/ViTAE-Transformer-Remote-Sensing/blob/main/Semantic%20Segmentation/mmseg/models/backbones/__init__.py

Thanks! It worked by importing the semantic segmentation project folder as mmsegmentation and uninstalling original mmsegmentation:

pip uninstall mmsegmentation
cd Semantic\ Segmentation
pip install -e .

I'm running from a notebook similar to demo/inference.ipynb, which i made.

from vitae-transformer-remote-sensing.

I'd succeeded on using the pre-trained model of ISPRS Potsdam.
Just a last request, could you also provide, please, the pre-trained model to the MillionAID, mentioned on the paper?

from vitae-transformer-remote-sensing.

Oh, sorry, it was on the root of the project, I didn't saw it. Thanks again! :D

from vitae-transformer-remote-sensing.

Sorry for bothering you, but a last question/request! 😅
The pre-trained model RSP-ViTAEv2-S-E100 on home page is trained for classification, isn't?
Do you have the RSP-ViTAEv2-S-E100, but trained for semantic segmentation on MillionAID dataset?

from vitae-transformer-remote-sensing.