leaplabthu / arc Goto Github PK

View Code? Open in Web Editor NEW

98.0 98.0 4.0 7.46 MB

ICCV 2023: Adaptive Rotated Convolution for Rotated Object Detection

License: Apache License 2.0

Python 91.07% Dockerfile 0.03% C++ 4.92% Cuda 3.90% C 0.03% Shell 0.06%

arc's People

Contributors

Stargazers

Watchers

Forkers

hunglq-vht cv-det hitmingzhiwang lkangelos

arc's Issues

配置环境的最后一步pip install -v -e . 实现不了

前面的步骤都是一样的，最后这个尤其是mmdet部分很难搞

The problem of train

Thanks for your impressive work!
When I use the command " python tools/train.py configs/obb/arc/arc_orcnn_r50fpn1x_ss_dota10_RxFFF_n4.py" for training，it will appear the error"File "/home/amax/anaconda3/envs/arc/lib/python3.7/site-packages/mmcv/utils/registry.py", line 44, in build_from_cfg f'{obj_type} is not in the {registry.name} registry')KeyError: 'OrientedRCNN is not in the models registry'"
Wonderng if it's a version issue with mmcv?Below is the version of the package I installed.

很感谢你的工作，想请教一下你最好的dota结果是：ARC-R50 81.77，test集是怎么进行测试的？

我自己复现的结果只有79点多，所以想给你确认下test的分割方法和提交策略，是不是哪里不同，导致了结果的差异性

关于测试时命令行中的save_dir

python tools/test.py configs/obb/arc/arc_orcnn_r50fpn1x_ss_dota10_RxFFF_n4.py
YOUR_CHECKPOINT_PATH --format-only --options save_dir=YOUR_SAVE_DIR;

作者您好，感谢您的开源代码，我在执行测试命令时，其中的--options后跟的save_dir配置，我不是很明白它的功能是什么，想请您解答一下。

在HRSC2016数据集上，用您的代码训练36轮后测试的精度是89.06(RTX3090, samples_per_gpu=8)，并没有达到90.41，请问是什么原因造成的，可以提供一下训练好的checkpoint吗

关于旋转矩阵的生成，有一些困惑

`def _get_rotation_matrix(thetas):
bs, g = thetas.shape
device = thetas.device
thetas = thetas.reshape(-1) # [bs, n] --> [bs x n]

x = torch.cos(thetas)
y = torch.sin(thetas)
x = x.unsqueeze(0).unsqueeze(0)  # shape = [1, 1, bs * g]
y = y.unsqueeze(0).unsqueeze(0)
a = x - y
b = x * y
c = x + y

rot_mat_positive = torch.cat((
    torch.cat((a, 1-a, torch.zeros(1, 7, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 1, bs*g, device=device), x-b, b, torch.zeros(1, 1, bs*g, device=device), 1-c+b, y-b, torch.zeros(1, 3, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 2, bs*g, device=device), a, torch.zeros(1, 2, bs*g, device=device), 1-a, torch.zeros(1, 3, bs*g, device=device)), dim=1),
    torch.cat((b, y-b, torch.zeros(1,1 , bs*g, device=device), x-b, 1-c+b, torch.zeros(1, 4, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 4, bs*g, device=device), torch.ones(1, 1, bs*g, device=device), torch.zeros(1, 4, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 4, bs*g, device=device), 1-c+b, x-b, torch.zeros(1, 1, bs*g, device=device), y-b, b), dim=1),
    torch.cat((torch.zeros(1, 3, bs*g, device=device), 1-a, torch.zeros(1, 2, bs*g, device=device), a, torch.zeros(1, 2, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 3, bs*g, device=device), y-b, 1-c+b, torch.zeros(1, 1, bs*g, device=device), b, x-b, torch.zeros(1, 1, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 7, bs*g, device=device), 1-a, a), dim=1)
), dim=0)  # shape = [k^2, k^2, bs*g]

rot_mat_negative = torch.cat((
    torch.cat((c, torch.zeros(1, 2, bs*g, device=device), 1-c, torch.zeros(1, 5, bs*g, device=device)), dim=1),
    torch.cat((-b, x+b, torch.zeros(1, 1, bs*g, device=device), b-y, 1-a-b, torch.zeros(1, 4, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 1, bs*g, device=device), 1-c, c, torch.zeros(1, 6, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 3, bs*g, device=device), x+b, 1-a-b, torch.zeros(1, 1, bs*g, device=device), -b, b-y, torch.zeros(1, 1, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 4, bs*g, device=device), torch.ones(1, 1, bs*g, device=device), torch.zeros(1, 4, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 1, bs*g, device=device), b-y, -b, torch.zeros(1, 1, bs*g, device=device), 1-a-b, x+b, torch.zeros(1, 3, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 6, bs*g, device=device), c, 1-c, torch.zeros(1, 1, bs*g, device=device)), dim=1),
    torch.cat((torch.zeros(1, 4, bs*g, device=device), 1-a-b, b-y, torch.zeros(1, 1, bs*g, device=device), x+b, -b), dim=1),
    torch.cat((torch.zeros(1, 5, bs*g, device=device), 1-c, torch.zeros(1, 2, bs*g, device=device), c), dim=1)
), dim=0)  # shape = [k^2, k^2, bs*g]

mask = (thetas >= 0).unsqueeze(0).unsqueeze(0)
mask = mask.float()                                                   # shape = [1, 1, bs*g]
rot_mat = mask * rot_mat_positive + (1 - mask) * rot_mat_negative     # shape = [k*k, k*k, bs*g]
rot_mat = rot_mat.permute(2, 0, 1)                                    # shape = [bs*g, k*k, k*k]
rot_mat = rot_mat.reshape(bs, g, rot_mat.shape[1], rot_mat.shape[2])  # shape = [bs, g, k*k, k*k]
return rot_mat`

在这个生成旋转矩阵的函数中，我不太理解得到rot_mat_positive 和rot_mat_negative这两个矩阵的拼接原理。这只适用于旋转3x3标准卷积，我希望理解这段代码以试图将其扩展到更大的卷积上。
能否提供一些理论上的支持，非常感谢

reresnet18

Can the ARC module be applied to REResNet-18?

pretrained backbone

I want to request how to use pretrained backbone?
At the test code,I can see the load_checkpoint to load the model.
But when training, can I how to use pretrained backbone?

DOTA Test

After get the example_Task1.zip.
I submit it on the website.http://bed4rs.net:8001/evaluation1/
but I don't get the result.
Who face this issue?how to solve?

mmpycocotools安装问题

请问如何正确安装mmpycocotools库，查询资料后说安装cpython=0.29.33，但依旧无法解决，希望有好心人可以帮忙解决，真心的感谢。

Which dataset is the pretrained arcresnet trained on?

Thanks for your work, I will appreciate for your reply.

关于Merging patch bboxes into full image阶段task速度变为0的问题

您好，我执行测试命令时，在Merging patch bboxes into full image阶段task处理速度变为0，如图所示。
我在github上搜索相关issues，有解决方案提到“在--options后面增加nproc=1或直接在代码中将nproc改为1，我看到ARC仓库的代码中已经有相关的修改了，实际执行的时候也是single processing。

(https://github.com/jbwang1997/OBBDetection/blob/master/mmdet/datasets/obb/dota.py#L76)

我的电脑显卡是3060 laptop版本，请问作者在执行的时候有碰到过这种问题吗？这种问题是否是因为电脑配置低问题引起的？
我观察到从task从7到9是在缓慢执行的，但是速度实在是太慢了，到9就基本上卡住很长时间。

Which dataset does the ARC-ResNet pretrain on? And if I wanna to do some modification，should I do pretrain again?

安装mmcv问题

当我的CUDA版本为12.2时似乎无法按照你的方法安装mmcv，有什么解决办法吗

The details of pretraining ARC

Thanks for your impressive work!
I wonder that how many epochs are needed to train Imagenet and then use the pretrain model to train the DOTA dataset? And is the results in Table. 3 under single-scale training ?

Question with how to replace the original resnet101 backbone to arcrenet101

hi, I would like to ask how I could replace the original resnet101 backbone with ur arcresnet101. The code from other papers uses pretrained restnet101 as the backbone. Should I just replace the pre-trained model with yours?

ImportError: /home/user1/SWC/ARC/mmdet/ops/corner_pool/corner_pool_ext.cpython-37m-x86_64-linux-gnu.so: file too short

when i already finish the install and go to train，i run zhe code “python tools/train.py configs/obb/arc/arc_orcnn_r50fpn1x_ss_dota10_RxFFF_n4.py;”
i meet such a question “ImportError: /home/user1/SWC/ARC/mmdet/ops/corner_pool/corner_pool_ext.cpython-37m-x86_64-linux-gnu.so: file too short”