lx-cly / yolov5_obb_kld Goto Github PK

KLD实现旋转目标检测

License: GNU General Public License v3.0

Dockerfile 0.31% Shell 0.63% Python 62.35% Makefile 0.01% C++ 30.80% Cython 0.13% Cuda 2.63% SWIG 0.03% Jupyter Notebook 3.12%

yolov5_obb_kld's Introduction

YOLOv5_OBB_KLD

代码实现了基于YOLOv5的遥感旋转框检测。利用CSL和KLD实现角度的学习，并且加入注意力机制提高检测效果。

数据集和权重文件

数据集 : 预处理过后的数据集.若是想训练自定义数据集，预处理过程参看项目.
原始预训练的权重文件 :
- yolov5x.pt、yolov5l.pt、yolov5m.pt、yolov5s.pt: Baidu Drive(6666).
训练好的部分权重文件 :
- YOLOv5_DOTAv1.5_OBB.pt: Baidu Drive(6666)

项目安装 (支持Linux系统)

1. Python 3.8 with all requirements.txt dependencies installed, including torch==1.6, opencv-python==4.1.2.30, To install run:

pip install -r requirements.txt

2. Install swig

cd  \.....\yolov5_OBB_KLD\utils
sudo apt-get install swig

3. Create the c++ extension for python

swig -c++ -python polyiou.i
python setup.py build_ext --inplace

训练

train.py. Note：修改.\models\yolo.py的Detect类中初始化函数的self.angle = 180 #CSL对应180 KLD对应1，默认使用CSL.

python train.py --weights weights/yolov5m.pt --cfg models/yolov5m.yaml --use_kld False --device 0 --epochs 300 --batch_size 4 --workers 4 --logdir runs/

评估

detect.py. Detect and visualize the detection result. Get the detection result txt.
evaluation.py. Merge the detection result and visualize it. Finally evaluate the detector

python detect.py --weights runs/exp/weights/best.pt --source 'dataset path' --output 'output path' --conf_thres 0.35 --iou_thres 0.4 --device 0 --kld False 
python evaluation.py 
''' example
检测结果已merge
检测结果已按照类别分类
校验数据集名称文件已生成
classname: ship
P: 0.8550878121966288
R: 0.900046446818393
[email protected]: 0.8889719225631516
classaps:  [     88.897]
原始存在文件,删除
检测结果已按照类别分类
校验数据集名称文件已生成
classname: ship
P: 0.8511538986754063
R: 0.8677432827509397
[email protected]: 0.8096364184338725
classaps:  [     80.964]
'''

结果展示

数据集图片尺寸裁剪为1024*1024，gap为10%。实验中NMS时统一使用的置信度阈值是0.35，IoU阈值是0.4。

感激

感谢以下的项目,排名不分先后

关于作者

  Name  : "lx"
  describe myself："good man"

yolov5_obb_kld's People

Contributors

Stargazers

Watchers

Forkers

yanxioa qixuxiang louderthanthunderx1 jie311 liulongee junhai0428 dmuqlzhang luo77123 jndxchengjiabao phd-haunglei

yolov5_obb_kld's Issues

question

请问你的yolov5s_decoupled head.yaml 里面说要增加分支请问应该怎么去增加分支来实现解耦头?
谢谢

作者你好，感谢你分享的工作，在阅读源码时有几个问题。
您给出的KLD损失绘图如下，显示分类损失一直为0，而角度损失一直下降。

但是，我在阅读compute_loss_KLD中，langle 初始值为0 ，

而后，langle 未参与其他计算，就到达最后损失计算，langle *= h['angle'] *S，所以 langle=0，即 langle 一直为0，角度仅在Box损失计算中参与Iou的计算。如下图

我试着训练DOTA数据集，效果不好。想知道您给的数据集label是什么格式（poly的4点坐标，还是由poly经过CV外接最小矩形的带角度的坐标，由OpenCV法转化为长边法的代码在哪里喃），谢谢
PS，这是我用CSL测全部16类后，得到的ship:0.8941434649109584, 但是现在KLD得到效果不好

权重失效了

权重链接失效了，没法下载了

Angular definition problem

My derivation of your repo is here: https://github.com/Suppersine/YOLOv7_obb_KFIOU

According to MMRotate, the CSL loss function abides by the "Long edge" angular definition, ranging from [-pi/, & pi/2, while the KLD & KFIOU loss functions abide by the "Old OpenCV (OOCV)" definition, ranging from [-pi/2, 0). To convert between these. I have written the function inside the rboxs_utils.py file below:

def regular_theta(theta, mode='180', start=-pi/2):
    """
    limit theta ∈ [-pi/2, pi/2)
    """
    assert mode in ['360', '180']
    cycle = 2 * pi if mode == '360' else pi

    theta = theta - start
    theta = theta % cycle
    return theta + start

def lebox2ocbox(x, y, w, h, theta):
    x, y, = x, y #ocbox[:2] = lebox[:2]
    #ocbox[-2:] = lebox[-2:]
    if theta < 0:
        return x, y, w, h, theta
    else:
        w, h = h, w
        theta -= pi/2
        return x, y, w, h, theta

def ocbox2lebox(x, y, w, h, theta):
    x, y, = x, y #lebox[:2] = ocbox[:2]
    #lebox[-2:] = ocbox[-2:]
    if w == max(w, h): 
        return x, y, w, h, theta
    else:
        w, h = h, w
        theta += pi/2
        return x, y, w, h, theta

and changed the "poly2rbox" by adding the highlighted section below to switch to OOCV angular definition:

In the loss file in the training phase under the KLD loss mode, however, when I printed out the predicted angles, they were supposed to be all-negative, but sometimes their max angles show positive values. (I guess I may not have changed their angular definition). I want to ask you... Since I can change the truth boxes' angular definition by the process above, where can I change the predicted boxes' angular definition?

训练过程

你好，请问你在训练过程中出现精度振荡范围大的情况吗

KLD训练得到权重后，在检测时发现角度信息相差不大

KLD训练得到权重后，在检测时发现角度信息相差不大，请问是标签或者是其他地方需要调整吗。
这是三个不同图片检测出来的结果，经过换算后，他们的角度信息几乎一致，检测结果看起来像是水平框检测
((876.5000610351562, 768.5), (462.3126525878906, 404.125), -87.8736572265625)
((862.5, 677.5000610351562), (275.181640625, 507.555908203125), -87.9666748046875)
((696.5, 1057.5001220703125), (185.2694854736328, 233.78030395507812), -87.05175018310547)

角度输出

作者你好！我想请问一下，想要在验证的时候保存目标框的旋转角度的话，需要在哪里改动呢？感谢！

大佬，训练DOTA1.5单类ship时为啥检测效果这么差啊

detect.py 检测效果只有0.02？大佬麻烦看下

detect报错

大佬大佬，请问要怎么样解决？
使用TPH训练，使用检测命令加载该训练权重时 python detect.py --weights runs_swin/exp/weights/best.pt 出现报错
Fusing layers...
Traceback (most recent call last):
File "detect.py", line 350, in
detect(kld_flag=opt.kld)
File "detect.py", line 54, in detect
model = attempt_load(weights, map_location=device) # load FP32 model
File "/content/drive//YOLOv5_OBB_KLD/models/experimental.py", line 150, in attempt_load
model.append(torch.load(w, map_location=map_location)['model'].float().fuse().eval()) # load FP32 model
File "/content/drive//YOLOv5_OBB_KLD/models/yolo_new.py", line 165, in fuse
if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'): # 如果函数层名为Conv标准卷积层，且同时层中包含‘bn’属性名
TypeError: isinstance() arg 2 must be a type or tuple of types