您是否可以给一个百度网盘的链接？onedrive下载 慢且不稳定。感谢

作者大大，我在把项目部署到本地的过程中报错了，但压根找不到报错的原因，求解FileNotFoundError: [Errno 2] No such file or directory: '/root/dw/pretrn/sam_vit_h_4b8939.pth'

作者大人，我反复翻看往期issue，始终不太明白用mmseg处理postsdam数据集的目的以及作用。我已经按照readme完成了模型的预训练工作，需要用postsdam数据集对模型进行finetune。参照往期issue，我选用3_Ortho_IRRG.zip和5_Labels_all.zip。我去除了3_Ortho_IRRG.zip中所有的tfw文件，只保留了tif格式。同样，5_Labels_all.zip中也只保存了tif格式文件。然后整个文件夹的结构如下图所示。我不明白还需要用mmseg处理什么内容，是用mmseg进行训练吗？还是对图像进行裁剪（因为图像尺寸6000*6000太大的原因）？还有为什么不直接使用裁剪好的RGB图像呢。（小人实在是弄得有点迷糊了，希望博主能援助一下，完成整个项目后，我想录制一个从0复现整个项目的视频，希望对后面学习的同学有所帮助）。

期待代码的发布，谢谢

博主，请问一下main_sam_hbox_semantic.py对dota数据集处理生成数据集，具体代码需要做什么修改。因为看到代码中默认的是对dior数据集的处理路径

[{'mask': {'size': [1024, 1024], 'counts': 'cRYf07fo04M3M2O2O01O1N2O0O3M2LbmZ9'}, 'bbox': array([714., 76., 726., 95.]), 'category': 'small-vehicle', 'size': 159, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'kRjf06fo05M2N3N10000O2O01N1O1N3L4M^mg8'}, 'bbox': array([731., 82., 744., 101.]), 'category': 'small-vehicle', 'size': 199, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'YWWd02lo04L4L3L4N1O2N1000O100O1N3N1N3N3LQiW;'}, 'bbox': array([648., 222., 667., 242.]), 'category': 'small-vehicle', 'size': 250, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'nSjf01no02N2N2O001N1000010O0010O001O10O01O1O100Onk8'}, 'bbox': array([733., 122., 754., 136.]), 'category': 'small-vehicle', 'size': 152, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': ']TQh02lo04M2O001O000010O0010O0010O001N3Mak]7'}, 'bbox': array([770., 137., 788., 151.]), 'category': 'small-vehicle', 'size': 136, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'lTUi06io02O1O001O00000001O0010O010OO101N2NQkX6'}, 'bbox': array([806., 153., 827., 168.]), 'category': 'small-vehicle', 'size': 153, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'UTce05io03L4M2O1O101O0000O2O0O2I9KQln9'}, 'bbox': array([691., 122., 705., 141.]), 'category': 'small-vehicle', 'size': 175, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'YTke02lo03M3L4M2O10000001O0O1O2N2KRlf9'}, 'bbox': array([700., 126., 714., 145.]), 'category': 'small-vehicle', 'size': 155, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'dTQf05io03J6N1O2N1000O101N1O2M2M4Mek9'}, 'bbox': array([706., 140., 718., 157.]), 'category': 'small-vehicle', 'size': 182, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'fSel05io0201O001O0010O010O01O01O010O001O1O0O2NVlf2'}, 'bbox': array([917., 117., 938., 131.]), 'category': 'small-vehicle', 'size': 166, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': ']Skl05jo02N101O1O000010O01O10O001O001O1O001O1O0O_l_2'}, 'bbox': array([924., 106., 945., 122.]), 'category': 'small-vehicle', 'size': 188, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'kRXm05jo02O001O001O010O00010O01O010O001N102LQmT2'}, 'bbox': array([935., 88., 955., 104.]), 'category': 'small-vehicle', 'size': 145, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'TYaj07go03L3N2N3N1O1O101O00O2O0O1M3N3M2N4KWgl4'}, 'bbox': array([849., 282., 864., 303.]), 'category': 'small-vehicle', 'size': 281, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'V[_h05io03M3M2N3N2O1N11O01O0001M3E[eR7'}, 'bbox': array([785., 349., 796., 367.]), 'category': 'small-vehicle', 'size': 212, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': '\[hh04jo03K5N1M3N2000O100O2M2M4MPej6'}, 'bbox': array([793., 354., 805., 370.]), 'category': 'small-vehicle', 'size': 168, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'c[i05ho04M2N2N2N3O00O100O1O1N3N2L5LfdP6'}, 'bbox': array([817., 361., 830., 376.]), 'category': 'small-vehicle', 'size': 188, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'WRVo05ho04M3M2N200O1000O1O2N1N4KSn<'}, 'bbox': array([ 999., 62., 1011., 80.]), 'category': 'small-vehicle', 'size': 167, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': ']Yec01no02M2HMPO6_o06010O01O01O0O2O0O2M2Nmfk;'}, 'bbox': array([631., 291., 646., 308.]), 'category': 'small-vehicle', 'size': 156, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': 'VY_c06ho04L2N3N100000000001O001N2O1M3M2O2Mmfn;'}, 'bbox': array([623., 287., 636., 304.]), 'category': 'small-vehicle', 'size': 216, 'label': 9}, {'mask': {'size': [1024, 1024], 'counts': '^Zmb02ko05K4N1N2N20000000O101M2N3N1Nmec<'}, 'bbox': array([607., 324., 621., 340.]), 'category': 'small-vehicle', 'size': 176, 'label': 9}]

Great work! I have two questions: 1. What exactly is 'counts'? 2. Are there any polygons for instance segmentation? Thank you

感谢您开源的代码，我希望运行test_gpu.py但是报错了，请问您有遇到这个 问题吗

(oneformer) lscsc@lscsc-System-Product-Name:SAMRS/Pretraining and Finetuning/End_to_End$ CUDA_VISIBLE_DEVICES=0 python test_gpu.py --backbone 'vit_b' --dataset 'potsdam' --ms 'False' --mode 'test' --resume /media/lscsc/nas/yihan/SegAN/SAMRS/
Pretraining and Finetuning/weight/vit_b_samrs_mae_clip_checkpoint-1599.pth --save_path /media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining an
d Finetuning/Encoder_Decoder/output
Traceback (most recent call last):
File "test_gpu.py", line 14, in
from models import SemsegFinetuneFramework
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/models.py", line 9, in
from backbone.intern_image import InternImage
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/backbone/intern_image.py", line 18, in
from .ops_dcnv3 import modules as opsm
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/backbone/ops_dcnv3/modules/init.py", line 7, in
from .dcnv3 import DCNv3, DCNv3_pytorch
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/backbone/ops_dcnv3/modules/dcnv3.py", line 16, in
from ..functions import DCNv3Function, dcnv3_core_pytorch
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/backbone/ops_dcnv3/functions/init.py", line 7, in
from .dcnv3_func import DCNv3Function, dcnv3_core_pytorch
File "/media/lscsc/nas/yihan/SegAN/SAMRS/Pretraining and Finetuning/End_to_End/backbone/ops_dcnv3/functions/dcnv3_func.py", line 16, in
import DCNv3
ModuleNotFoundError: No module named 'DCNv3'

首先感谢您的工作！我在阅读SAMRS与RSP（An Empirical Study of Remote Sensing Pretraining）时，发现两篇文章公布的rsp-r50分割结果有所差异。
具体而言:
(1) SAMRS中表3 rsp-r50在potsdam结果为 OA=90.49 mF1=90.97

RSP中表6 rsp-r50的结果为OA=90.61 mF1=89.94

(2) SAMRS中表4 rsp-r50在isaid结果为mIoU为32.97

RSP中表7 rsp-r50的结果为mIoU为61.6

我不太清楚造成这种差异的原因是什么？特别是在isaid上的结果差异比较大，我有点不太确定该以哪篇文章的结果为准。期待您的回复。

作者大人，我在运行encoder_and_decoder下的main_finetune.py文件报错日志显示脚本 main_finetune.py 无法识别 --load 参数。查看main_finetune.py 也没有找到对参数load的编译代码。

I wanna ask how to generate custom datasets using the code or fine-tune the model on custom datasets using the code

When I open the onedrive link, I got this page information.

If possible, can you update your dataset download link. plz
thx

Not really an issue, more of a PSA:
For anyone downloading the data through the OneDrive link, I have found that if you go into the dataset folders and download each of those files individually (e.g inside the SIOR directory download the isinlabels.zip, samlabels.zip, test_images.zipm trainval_images.zip, train.txt and val.txt files) you get much higher speeds, for me it was < 1MB/s before up to a peak of 21MB/s downloading individually and the zips unpack correctly. I had issues unpacking the SIOR.zip file, but not downloading them individually.

@DotWang you could possibly add this recommendation as a note in the readme.

作者大人，我在运行encoder_and_decoder下的main_pretrain.py文件时候遇到了这个报错，显示在 datasets.py 文件的第 91 行代码中，尝试访问 self.args.background 时出现了 AttributeError，因为 background 属性没有定义在 args 对象中。但是运行main_pretrain.py文件的命令中没有background参数的设置。不明白是哪里出现了问题，请教博主大人

作者大人，我在网上下载的数据集结构与你的不太一样，导致预训练无法正常进行。请问作者大人可以把您使用的这些数据集的原链接给一下吗，或者把数据集上传。十分感谢！万分感谢

Congrats to your excellent work first!
I want to know if there is any plan to release the pre-training and finetuning codes?

Hi,

Thanks for the great work! I was wondering if you can provide the SAM hyperparmeter settings used to generate the dataset.

it's a good job，and I saw in your paper that you obtained the segmentation results of DOTA 2.0 version. Can I use this code to obtain the segmentation results of DOTA 1.0?
look forward to you reply,
thank you very much

作者大大，您好，最近在尝试复现您的论文，遇到一些问题请教一下
1.https://github.com/ViTAE-Transformer/SAMRS/tree/main/Pretraining%20and%20Finetuning 这个链接下的提供的Segmentation Pretrained Models提供的是直接可以用来测试ISPRS Potsdam数据集（我把图裁剪成了512）的吗？
2.我使用resnet50_upernet_imp_sep_model.pth这个权重测试ISPRS Potsdam数据集的时候发现测试出来的精度很低。
使用的指令是：
CUDA_VISIBLE_DEVICES=0 python Pretraining_and_Finetuning/Encoder_Decoder/test_gpu.py --backbone 'resnet50' --decoder 'upernet' --dataset 'potsdam' --ms 'False' --mode 'test' --resume resnet50_upernet_imp_sep_model.pth --save_path ./save

加载权重时候会出现这个，不知道这会不会影响权重的正确载入？
最后日志文件是：

请问您觉得有什么可能发生的问题会影响精度测试，我排查一下
3.请问可以提供百度云链接的数据集下载方式吗？onedrive下载会一直中断

请问作者有没有在遥感数据集上微调好的权重呢

请问作者，数据集的H, W必须相同吗，如果是高宽不一致的图像该如何得到对应( H, W )的mask

cuda12.2 在conda里面配置相关环境一直不成功，需要对应什么版本呀

Hello,

After going through your data, you just labeled objects which has boxes. The background like sky or water are not labeled.

Therefore, I am curious how your data can be used for semantic segmentation as you claimed in your conclusion on page 6.
Thanks,

如图所述，Encoder Decoder可以跑通，End to End不行

What source code should I run to create an automatic label dataset?

Hello,

We are wondering if a picture has multiple objects, you just use gray labels. Therefore, i am curious how i can distinguish different objects in a picture. Is it possible to share codes to generate pictures like your Figure 6 in your SAMRS paper?

Thanks,
Liya

Is the dataset copyright free of charge?

I'm trying to participate in a certain competition, and I'm asking because I'm trying to learn with that dataset.

Thanks

您好，请问作者您有将SAM直接在遥感大数据集上微调的模型吗

请问各位遇到这个DCNv3模块问题怎么解决的呀。（研0新手小白）

你好，想请教下您生成的SAMRS中数据集实例分割标注的格式是怎样的，以SIOR为例解析ins标注文件夹下的pkl文件得到如下信息：

我本身是想用实例分割的mask标注进行可视化，但是这里的mask标签是一串字符串，是我解析的方式有什么问题吗？

作者大人，我在预训练时候根据您github上的示例命令选择--decoder 'upernet'。但报错显示SemsegPretrnFramework 类中没有对‘upernet’ decoder 的处理逻辑，我翻看了model.py文件也没有找到。所以想问一下作者大人。以下是我自己的运行命令：CUDA_VISIBLE_DEVICES=0 python main_pretrain.py --backbone 'resnet50' --decoder 'upernet' --datasets 'sota' 'sior' 'fast' --batch_size 12 --batch_size_val 12 --workers 8 --save_path '/root/autodl-tmp/SAMRs/save' --distributed 'False' --end_iter 80000 --image_size 224 --init_backbone 'imp'

作者大人，我在运行main_finetune.py时候遇到报错。具体是main_finetune.py表明在计算 F1 分数时出现了除零错误。我怀疑是我postdam数据集的结构有问题吗？下图是我的文件夹结构。其中images文件夹下保存rbg图像（格式jpg），train.txt和valid.txt分别保存训练集和验证集图片的名称，lables文件夹下保存images文件夹中相同图像的分割图像。

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --nproc_per_node=8
--nnodes=1 --master_port=10001 --master_addr = [server ip] main_pretrain.py
--backbone 'resnet50' --decoder 'upernet'
--datasets 'sota' 'sior' 'fast'
--batch_size 12 --batch_size_val 12 --workers 8
--save_path '[SEP model save path]'
--distributed 'True' --end_iter 80000
--image_size 224 --init_backbone 'imp'
作者大人，这个分布式预训练运行脚本是一机多卡式还是多机多卡式的。如果我想只用一个gpu单卡运行，可以吗？需要对main_pretrain.py文件进行修改吗？

Thank you very much for sharing your research code.

I have one question, and I hope you can help me.

In the files main_finetune.py, main_pretrain.py, and test_gpu, the datasets are set to POTSDAM, VAIHNGEN, and ISAID, instead of the FAST, SOTA, and SIOR datasets that you uploaded to OneDrive. Should I modify the code to use the FAST, SOTA, and SIOR datasets instead?

Looking forward to your response.
Thank you so much.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --nproc_per_node=8
--nnodes=1 --master_port=10001 --master_addr = [server ip] main_pretrain.py
--backbone 'resnet50' --decoder 'upernet'
--datasets 'sota' 'sior' 'fast'
--batch_size 12 --batch_size_val 12 --workers 8
--save_path '[SEP model save path]'
--distributed 'True' --end_iter 80000
--image_size 224 --init_backbone 'imp'
作者大人，这个分布式预训练运行脚本是一机多卡式还是多机多卡式的。如果我想只用一个gpu单卡运行，可以吗？需要对main_pretrain.py文件进行修改吗？