This is the data repository for CBPH-Net, and further usage details will be provided after the acceptance of the paper. The dataset STBD-08, created in this study, is available for download. However, due to privacy concerns related to the research topic, please email the corresponding author of the paper to obtain a copy of the agreement. This agreement ensures that the dataset will not be used for unintended academic purposes.
I am sorry for the prevoious error code, it has some problems, reasons and solution you can see issue #2. After you download the code，please delete model/CBPH.py, it has been rewritten into model/common.py, but you can still read it to learn. I update a new version, you can just run it by "python train.py --cfg models/CBPH-Net.yaml --data data/voc_bm.yaml --batch-size 64 --epoch 100".
For the dataset, please consult "https://github.com/Whiffe/SCB-dataset", it has more related datasets and i have sent my dataset to this researcher.
之前的代码有一些小错误，详情和解决请见我在issues #2里的回答。我对于之前存在的错误十分抱歉。
我上传了新的代码，并确保其可以运行。重新注明了如何安装CBPH头，如果您不想使用，注释掉代码中对应的部分，并在yaml文件中将CBPH更换为Detect即可。
如果您想关闭其中的模块，请自行注释掉不需要的部分就好。代码仓库是基于YOLOv5的，网上有很多教程，我的代码基本上也是在其基础上进行修改的（在这里很抱歉地回复某个咨询者，我没能力自行创作全新的模型，只能在这里修修改改，如果有我就会去发顶会而不是2区不知名期刊）。
关于数据集，请移步https://github.com/Whiffe/SCB-dataset，这里有更丰富的数据集资源，之前的数据集我也授权给这位老师了。 今后不会再回复咨询邮件，因为我将于2024/5月从华中师范大学毕业，教育邮箱也会被收回，未来也无意在这个领域进行深耕，请勿打扰。
github clone https://github.com/icedle/CBPH-Net`
cd CBPH-Net
pip install -r requirements
The default detection head used is the one from YOLOv5. However, we have integrated CBPH into the project (refer to models/CBPH.py). If you want to enable CBPH as the detection head, please follow the steps below:
1. Go to yolo.py, find isinstance (m, detect), and add the following code after it:
if isinstance(m, CBPH):
s = 256 # 2x min stride
m.inplace = self.inplace
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward
m.anchors /= m.stride.view(-1, 1, 1)
check_anchor_order(m)
self.stride = m.stride
2. Go to models/yolo.py, find:
def _profile_one_layer(self, m, x, dt):
c = isinstance(m, Detect):
and modify it as:
c = isinstance(m, Detect, CBPH):
3. Go to models/yolo.py, find:
def _apply(self, fn)
if isinstance(m, Detect):
and modify it as:
c = isinstance(m, Detect) or isinstance (m, CBPH):
4. Go to models/yolo.py, find "parse_model" function, and add the following code:
elif m is CBPH:
args.append([ch[x] for x in f])
if isinstance(args[1], int): # number of anchors
args[1] = [list(range(args[1] * 2))] * len(f)
5. Go to models/CBPH-Net.yaml
Replace the "Detect" in the last line with "CBPH".
We give an example, more details please see train.py
python trian.py --cfg models/CBPH-Net.yaml --batch-size 32 --epoch 100
Please see val.py
We give an example
The dataset format is VOC (.xml). You can use to convert it to COCO(.json) or TXT format. We already provide a script to transfer xml format to txt format (Please see prepare_data.py).
