bosszard/rotation-yolov5

rotation detection based on yolov5

the rotation detection

Requirement

torch==1.6
shapely==1.7.1
opencv==4.2.0.34

inference

you can download the weights BaiduYun(password is 4ud5) or GoogleDrive for ship detection by my dataset(not DOTA) to test the demo.

$ python detect.py

train

what format my model need

Not much different from yolo dataset,just add an angle and we define the box attribute w is always longer than h!

So wo define the box label is (cls, c_x, c_y, Longest side，short side, angle)

Attention!we define angle is a classify question,so we define 180 classes for angle.

For Example: Range for angle is [-90，90), so wo should add 90 in angle while make your dataset label and then your label's Range should be [0,179)

modify yaml

models/yolov5m.yaml: set nc to your dataset class num; data/wheat0.yaml:set nc to your dataset class num, and set names to your dataset class name;

$ python train.py

update

2021.1.4---correct some BUG for training

details

If you have any question,welcome discuss with me by This or email to [email protected]

rotation-yolov5's Issues

RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity

I am running the algorithm on a Mac OS and this is the error I get when I run the detect.py file

run code:
python detect.py --weights yolov5s.pt --source data/images/carpark1.jpg

error output:
Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.4, device='', img_size=1024, iou_thres=0.5, output='inference/output', save_txt=False, source='data/images/carpark1.jpg', update=False, view_img=False, weights=['yolov5s.pt'])
Using CPU

Fusing layers...
Model Summary: 140 layers, 7.45958e+06 parameters, 0 gradients
image 1/1 /Users/syazaniwahid/Documents/UROPS/YOLO/YOLOV5_RBB/rotation-yolov5/data/images/carpark1.jpg: Traceback (most recent call last):
File "detect.py", line 174, in
detect()
File "detect.py", line 79, in detect
pred = rotate_non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=True)
File "/Users/syazaniwahid/Documents/UROPS/YOLO/YOLOV5_RBB/rotation-yolov5/utils/general.py", line 845, in rotate_non_max_suppression
conf, j = x[:, 5:5 + nc].max(1, keepdim=True)
RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity

dota v1.5数据集训出来效果很差

           Class      Images     Targets           P           R      [email protected]  [email protected]:.95: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 115/115 [00:23<00:00,  4.92it/s]
             all         458    6.96e+04      0.0863       0.133       0.098      0.0347
           plane         458    2.55e+03       0.423       0.551       0.435       0.168
baseball-diamond         458         213       0.292       0.634       0.494       0.164
          bridge         458         466       0.114       0.337       0.157      0.0381

ground-track-field 458 145 0.225 0.407 0.324 0.128
small-vehicle 458 4.33e+04 0.327 0.198 0.159 0.0563
large-vehicle 458 5.14e+03 0 0 0 0
ship 458 1.08e+04 0 0 0 0
tennis-court 458 763 0 0 0 0
basketball-court 458 143 0 0 0 0
storage-tank 458 2.94e+03 0 0 0 0
soccer-ball-field 458 149 0 0 0 0
roundabout 458 185 0 0 0 0
harbor 458 2.1e+03 0 0 0 0
swimming-pool 458 576 0 0 0 0
helicopter 458 78 0 0 0 0
container-crane 458 14 0 0 0 0
基本上在val集上，没有识别出来，16个类
我是在train上训练了500epoch，不应该有问题啊，求解答。

dota数据集测试耗时问题

你在做detect时，在dota验证和测试集上一般用时要多少呢？有没有做测试加速呢？

detector.py问题

你好，你的这个项目很有参考价值。我使用了你的旋转框yolov5，我下载了你的权重文件，当我运行detect.py文件时，存在以下问题，请问怎么解决呢
/home/miki/anaconda3/envs/yolov5/bin/python /home/miki/project/object_detection/rotation-yolov5/detect.py
Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.4, device='', img_size=1024, iou_thres=0.5, output='inference/output', save_txt=False, source='inference/images', update=False, view_img=False, weights='weights/m-70.pt')
Using CUDA device0 _CudaDeviceProperties(name='GeForce GTX 1050 Ti', total_memory=4036MB)

Fusing layers...
Model Summary: 188 layers, 2.22122e+07 parameters, 0 gradients
image 1/1 /home/miki/project/object_detection/rotation-yolov5/inference/images/1.tif: Traceback (most recent call last):
File "/home/miki/project/object_detection/rotation-yolov5/detect.py", line 174, in
detect()
File "/home/miki/project/object_detection/rotation-yolov5/detect.py", line 79, in detect
pred = rotate_non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=True)
File "/home/miki/project/object_detection/rotation-yolov5/utils/general.py", line 869, in rotate_non_max_suppression
i = nms(boxes, scores)
File "/home/miki/project/object_detection/rotation-yolov5/utils/general.py", line 798, in nms
rotate_ious = rotate_iou(max_iou_boxes[-1], sort_boxes[1:])
File "/home/miki/project/object_detection/rotation-yolov5/utils/general.py", line 710, in rotate_iou
r_b1 = get_rotated_coors(box1[i])
File "/home/miki/project/object_detection/rotation-yolov5/utils/general.py", line 729, in get_rotated_coors
R[:2] = cv2.getRotationMatrix2D(angle=-a*180/math.pi, center=(cx,cy), scale=1)
TypeError: Argument 'angle' can not be treated as a double

Process finished with exit code 1
希望能得到你的回复，谢谢！

angle dataset

Hello. If my dataset have angle from -180 to 180, Can I use this repo?

dota test

Thanks for your sharing. I wonder what is your test result on the DOTA data set?

请问标注工具用哪个？

我尝试了https://github.com/chinakook/labelImg2，只能拿到angle，好像和文中的尺寸也对不上

不同

作者，您好！我对您的工作非常感兴趣，我有些疑问：rotation-yolov5与yolov5模型有什么不同，修改哪一部分？

运行detect.py时遇到了问题，这该怎么解决呢

Fusing layers...
Model Summary: 188 layers, 2.22122e+07 parameters, 0 gradients
image 1/1 D:\Weld inspection\YOLOv5\rotation-yolov5-master\inference\images\1.tif: Traceback (most recent call last):
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\detect.py", line 174, in
detect()
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\detect.py", line 79, in detect
pred = rotate_non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=True)
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\utils\general.py", line 869, in rotate_non_max_suppression
i = nms(boxes, scores)
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\utils\general.py", line 797, in nms
rotate_ious = rotate_iou(max_iou_boxes[-1], sort_boxes[1:])
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\utils\general.py", line 710, in rotate_iou
r_b1 = get_rotated_coors(box1[i])
File "D:\Weld inspection\YOLOv5\rotation-yolov5-master\utils\general.py", line 729, in get_rotated_coors
R[:2] = cv2.getRotationMatrix2D(angle=-a*180/math.pi, center=(cx,cy), scale=1)
TypeError: Can't parse 'center'. Sequence item with index 0 has a wrong type

数据集制作

你好，我想问一下，标签文件里 cx,cy,w,h 是怎么算出来的，网上的数据标签处理只是针对无旋转的YOLOv5，现在我训练出来的结果很差，，基本都不能看，所以我怀疑是数据集制作的问题

Weights not available anymore

Hello,

When I clicked the download link you've given for weights, it says "Page does not exist". Please send me the '10.28-0.5angle.pt' file.
Thank you.

数据集格式问题

您好，我自己制作数据集后，在训练时在wheat0.yaml中要按照什么样的方式进行排放呢？

预测问题

你好请问一下我用hrsc2016数据集训练完之后预测得到的box有的变成一条直线是怎么回事，就像照片中这种情况

retanglelabel2mylabel.py

请问 retanglelabel2mylabel.py，w，h是不是写反了？还有原始图片是不是必须得是长方形？

数据集制作

你好，我想问一下，标签文件里 cx,cy,w,h 是怎么算出来的，网上的数据标签处理只是针对无旋转的YOLOv5，现在我训练出来的结果很差，，基本都不能看，所以我怀疑是数据集制作的问题

预训练有关事宜

你好，我在查看了你的代码之后，发现你的代码似乎不适合用作预训练使用
rotation-yolov5中的最后一维是[x,y,w,h,class_num(5),angle(180)]
如果我的class_num n<5，我的输出变为[x,y,w,h,class_num(n),angle(180)]那么前4+n维都是正确的,用于预测xywh和class，但是后面[4+n:4+n+180]将使用原来用于预测类和角度的维度作为预训练，而不是使用原来的后180个维度作为预训练，似乎不符合预训练的原理。
所以应该将代码修改为[x,y,w,h,angle(180),class_num]
不知道我的理解是否正确，如果正确的话，我想将COCO数据集进行旋转之后获得一个合适的数据集，将rotation-yolov5在上面训练得到一个预训练模型，然后再在我的数据集上进行训练，这个想法是正确的吗？

Can't get attribute 'SiLU'

Traceback (most recent call last):
File "train.py", line 456, in
train(hyp, opt, device, tb_writer)
File "train.py", line 69, in train
ckpt = torch.load(weights, map_location=device) # load checkpoint
File "/home/yuanweiqiao/anaconda3/envs/CL/lib/python3.7/site-packages/torch/serialization.py", line 584, in load
return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
File "/home/yuanweiqiao/anaconda3/envs/CL/lib/python3.7/site-packages/torch/serialization.py", line 842, in _load
result = unpickler.load()
AttributeError: Can't get attribute 'SiLU' on <module 'torch.nn.modules.activation' from '/home/yuanweiqiao/anaconda3/envs/CL/lib/python3.7/site-packages/torch/nn/modules/activation.py'>

数据增强中上下翻转角度不变对么？

我的感觉是上下翻转和左右翻转角度都应该变成补角呢？请赐教

Yolov5 version

Hello. I use your repo then see my dataloader image with labels not corresponding to each other? Can you re-check your repo?
Thanks very much

一个BUG？

rotation-yolov5/utils/general.py

Line 533 in f68d7cc

ttheta = torch.zeros_like(ps[:, 10:])

这里设置了ttheta的形状，应该是根据后面的 t = torch.full_like(ps[:, 5:10], cn, device=device)来设置的
这里的t是设置为跟cls形状相同，但是默认固定为原始的5个类别，所以如果只是修改了yaml里的类别为其他，而不修改这里就会报错，这里是不是要修改一下弄成跟yaml联系才行
我在打算把这个模型用在我的自定义数据集上发现了这个问题，我是只有一个类别要识别的，我是最近才接触的目标检测，如果是我理解错了，希望能提出纠正一下。

test,py

Can I using test.py in your repo to test with rotated bounding box ?

train.py 运行到第二轮报运行时错误

...很多如下错误
/opt/conda/conda-bld/pytorch_1603729062494/work/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [60,0,0], thread: [60,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.
/opt/conda/conda-bld/pytorch_1603729062494/work/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [60,0,0], thread: [61,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.
/opt/conda/conda-bld/pytorch_1603729062494/work/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [60,0,0], thread: [62,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.
/opt/conda/conda-bld/pytorch_1603729062494/work/aten/src/ATen/native/cuda/IndexKernel.cu:84: operator(): block: [60,0,0], thread: [63,0,0] Assertion index >= -sizes[i] && index < sizes[i] && "index out of bounds" failed.
2/59 2.59G 0.0773 0.04795 0.05063 0.2194 0.3953 1 1024: 93%|█████▌| 12214/13191 [1:37:05<07:46, 2.10it/s]
Traceback (most recent call last):
File "train.py", line 456, in
train(hyp, opt, device, tb_writer)
File "train.py", line 269, in train
loss, loss_items = compute_loss(pred, targets.to(device), model) # loss scaled by batch_size
File "/home/cao/Downloads/rotation-yolov5-master/utils/general.py", line 530, in compute_loss
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
RuntimeError: CUDA error: device-side assert triggered

预测的问题

请问怎么在预测的时候显示概率和类别呢？

cv2.minAreaRect角度范围为[0,90]，左闭右闭

rotation-yolov5/retanglelabel2mylabel.py

Line 34 in 4122b52

rect = cv2.minAreaRect(data) # 得到最小外接矩形的（中心(x,y), (宽,高), 旋转角度）

https://blog.csdn.net/LEILEI18A/article/details/115006688
这是我的测试结果，没有小于0度的，也许是opencv修改了角度定义了吧

opencv版本原因导致的

problem while downloading weights I 下载权重时出现问题

Hi, I am trying to download the trained weights which you have added here. Could you help, to get them, please.

嗨，我正在尝试下载您在此处添加的经过训练的砝码。 您能帮忙，请得到他们。

PS: I don't understand Chinese)

角度问题

请问下你使用分类的方法求目标的角度吗？0-180°对应180个类别

AttributeError: 'LineString' object has no attribute 'exterior'

请问这个是什么原因呢

Traceback (most recent call last):
  File "detect.py", line 175, in <module>
    detect()
  File "detect.py", line 116, in detect
    plot_one_box(xywh, im0, label=label, color=colors[int(cls)], line_thickness=3, path=path)
  File "D:\master\code\car_detection\rotation-yolov5-master\utils\general.py", line 1332, in plot_one_box
    _x, _y = poly_box.intersection(maxbox).exterior.coords.xy
AttributeError: 'LineString' object has no attribute 'exterior'

图片形状问题

你好，请问图片必须是正方形1024*1024吗？可以是长方形的图片吗？

how i convert my image masks convert into rotated yolo format

I tried minimum area rectangle for find angle but they didn't work it.

BUG?

rotation-yolov5/utils/general.py

Line 533 in 5544637

ttheta = torch.zeros_like(ps[:, 5+class_num:])

我进行DEBUG的时候，发现compute_loss里的p的shape是[batch,3,128,128,5+180+class]
但是yolov5里的是[batch,3,128,128,4+class]，所以shape最后一个维度应该是4(x,y,w,h)+1(confidence)+1(angle)+class才对啊，这里的180是哪里来的，这样angle就是有180个维度了，但是应该是只有一个才对啊
输出的p应该是[batch,3,128,128,4+1+1+class]
并且ttheta = torch.zeros_like(ps[:, 5])才对吧 而不是ttheta = torch.zeros_like(ps[:, 5])

the function of def rectlong2opencv(boxes)

可以解释一下这个函数吗？

def rectlong2opencv(boxes):

    boxes_ = boxes.copy()
    boxes_[boxes[..., 4] < 0, 2] = boxes[boxes[..., 4] < 0, 3]
    boxes_[boxes[..., 4] < 0, 3] = boxes[boxes[..., 4] < 0, 2]
    boxes_[boxes[..., 4] >= 0, 4] = boxes[boxes[..., 4] >= 0, 4] - 90
    gt_box = boxes_.copy()
    gt_box[..., 2] = boxes_[..., 3]
    gt_box[..., 3] = boxes_[..., 2]
    return gt_box

boxes_[boxes[..., 4] < 0, 2] = boxes[boxes[..., 4] < 0, 3]
boxes_[boxes[..., 4] < 0, 3] = boxes[boxes[..., 4] < 0, 2]
这里第三四个参数互换，然后下面又换了过来
gt_box[..., 2] = boxes_[..., 3]
gt_box[..., 3] = boxes_[..., 2]

数据标签制作问题

请问在标签文件中的cx, cy是指中心点坐标还是左上角坐标，角度在制作的过程中做了什么归一化操作吗

about torch version and requirements

Hi，我在anachonda3中完成环境配置，尝试运行detect.py时遇到了以下问题，这些问题可能来源于torch版本或者其他的冲突
CUDA=10.0, cudnn=7.6.5, python=3.7, opencv-python=4.2.0.34
测试环境1：torch=1.6, torchvision=0.7, 其他包均从requirement.txt文件安装，运行detect.py报错：
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

测试环境2：torch=1.7, torchvision=0.8.1, 其他包均遵从U版yolov5 requirement.txt，降低opencv-python版本至4.2。i.e.标准yolov5环境，并测试通过yolov5。加装shapely。运行detect.py报错：
torch.nn.modules.module.ModuleAttributeError: 'Hardswish' object has no attribute 'inplace'
根据网上说法，这个问题需将torch降低至1.6，所以该问题死循环了。

the definition of angle?

the longside clockwise to x-axis is positive angle?
the final angle is the longside to x-axis and then add 90?
so the final angle of the three example you gave is 0+90, 40+90, -90+90, is this right?

AttributeError: 'Hardswish' object has no attribute 'inplace'

(yolov5rbb) C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5>python detect.py --weights yolov5s.pt --source data/images/carpark2.jpg
Namespace(weights=['yolov5s.pt'], source='data/images/carpark2.jpg', output='inference/output', img_size=1024, conf_thres=0.4, iou_thres=0.5, device='', view_img=False, save_txt=False, classes=None, agnostic_nms=False, augment=False, update=False)
Using CPU

Fusing layers...
Model Summary: 140 layers, 7.45958e+06 parameters, 0 gradients
image 1/1 C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\data\images\carpark2.jpg: Traceback (most recent call last):
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\detect.py", line 174, in
detect()
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\detect.py", line 75, in detect
pred = model(img, augment=opt.augment)[0]
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\models\yolo.py", line 116, in forward
return self.forward_once(x, profile) # single-scale inference, train
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\models\yolo.py", line 136, in forward_once
x = m(x) # run
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\models\common.py", line 89, in forward
return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\PC\Documents\UROPS\YOLOV5_RBB\rotation-yolov5\models\common.py", line 33, in fuseforward
return self.act(self.conv(x))
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\activation.py", line 472, in forward
return F.hardswish(input, self.inplace)
File "C:\Users\PC\anaconda3\envs\yolov5rbb\lib\site-packages\torch\nn\modules\module.py", line 1207, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'Hardswish' object has no attribute 'inplace'

How do I resolve the hardwish error