pytorch mmrotate训练自定义数据集xml转txt

摘要

mmrotate是旋转目标检测框架，效果如下图所示，主要讲解如果xml文件标注信息转成dota格式用txt存储标注信息，以及如何实现顺时针逆时针矩形框旋转。

数据处理

首先学会矩形框旋转运算，主要是利用sin和cos进行计算旋转后的x,y坐标点信息
逆时针公式

顺时针公式

data1.0数据格式

data
	train
		images
			1.png
			2.png
			3.png
		labels
			1.txt
			2.txt
			3.txt
具体txt格式为

具体代码案例

import math
x=1
y=1
angle=90
x1=x*math.cos(math.radians(angle))-y*math.sin(math.radians(angle));
y1=y*math.cos(math.radians(angle))+x*math.sin(math.radians(angle));
print((x1,y1))

逆时针旋转90度刚好为(-1,1)，顺时针换个公式就可以实现了。

自定义数据转dota格式

数据分为二个部分，一部分xml文件，另一部分图像

修改路径就可以使用了，我这里是逆时针代标注的代码，如果顺时针需要修改一下

import os
import xml.etree.ElementTree as ET
import math
import cv2 as cv
all_name=[]
change=['light truck','heavy truck']
change_dict={'light truck':'light-truck','heavy truck':'heavy-truck'}
def voc_to_dota(xml_path, xml_name):
    txt_name = xml_name[:-4] + '.txt'
    txt_path = xml_path + '/txt_label'
    if not os.path.exists(txt_path):
        os.makedirs(txt_path)
    txt_file = os.path.join(txt_path, txt_name)
    file_path = os.path.join(xml_path, file_list[i])
    tree = ET.parse(os.path.join(file_path))
    root = tree.getroot()
    # print(root[6][0].text)
    image_path = 'D:/racedata\detrota/2_Aligned_RGB_infrared\data/train\image-H/'
    out_path = 'D:/racedata\detrota/2_Aligned_RGB_infrared/data/show/'
    filename = image_path + xml_name[:-4] + '.png'
    img = cv.imread(filename)
    print(filename)

    with open('D:/racedata\detrota/2_Aligned_RGB_infrared\data/train/anno-H-txt-new/'+txt_name, "w+", encoding='UTF-8') as out_file:
        for obj in root.findall('object'):
            name = obj.find('name').text
            if name not in all_name:
                all_name.append(name)
            if name in change:
                name=change_dict[name]
            difficult = obj.find('difficult').text
            # print(name, difficult)
            robndbox = obj.find('rbox')
            cx = float(robndbox.find('cx').text)
            cy = float(robndbox.find('cy').text)
            w = float(robndbox.find('w').text)
            h = float(robndbox.find('h').text)
            angle = float(robndbox.find('angle').text)
            #angle=angle+90
            #angle = 180-angle
            # print(cx, cy, w, h, angle)
            p0x, p0y = rotatePoint(cx, cy, cx - w / 2, cy - h / 2, -angle)
            p1x, p1y = rotatePoint(cx, cy, cx + w / 2, cy - h / 2, -angle)
            p2x, p2y = rotatePoint(cx, cy, cx + w / 2, cy + h / 2, -angle)
            p3x, p3y = rotatePoint(cx, cy, cx - w / 2, cy + h / 2, -angle)

            # 找最左上角的点
            dict = {p0y:p0x, p1y:p1x, p2y:p2x, p3y:p3x }
            list = find_topLeftPopint(dict)
            #print((list))
            if list[0] == p0x:
                list_xy = [p0x, p0y, p1x, p1y, p2x, p2y, p3x, p3y]
            elif list[0] == p1x:
                list_xy = [p1x, p1y, p2x, p2y, p3x, p3y, p0x, p0y]
            elif list[0] == p2x:
                list_xy = [p2x, p2y, p3x, p3y, p0x, p0y, p1x, p1y]
            else:
                list_xy = [p3x, p3y, p0x, p0y, p1x, p1y, p2x, p2y]

            # 在原图上画矩形 看是否转换正确
            cv.line(img, (int(list_xy[0]), int(list_xy[1])), (int(list_xy[2]), int(list_xy[3])), color=(255, 0, 0), thickness= 3)
            cv.line(img, (int(list_xy[2]), int(list_xy[3])), (int(list_xy[4]), int(list_xy[5])), color=(0, 255, 0), thickness= 3)
            cv.line(img, (int(list_xy[4]), int(list_xy[5])), (int(list_xy[6]), int(list_xy[7])), color=(0, 0, 255), thickness= 2)
            cv.line(img, (int(list_xy[6]), int(list_xy[7])), (int(list_xy[0]), int(list_xy[1])), color=(255, 255, 0), thickness= 2)
            #data= str(str(cx)+" "+str(cy)+' '+str(w)+' '+str(h)+' '+str(angle))
            data = str(list_xy[0]) + " " + str(list_xy[1]) + " " + str(list_xy[2]) + " " + str(list_xy[3]) + " " + \
                    str(list_xy[4]) + " " + str(list_xy[5]) + " " + str(list_xy[6]) + " " + str(list_xy[7]) + " "
            data = data + name + " " + difficult + "\n"
            out_file.write(data)
        #cv.imwrite(out_path + xml_name[:-4] + '.png', img)

def find_topLeftPopint(dict):
    dict_keys = sorted(dict.keys())  # y值
    temp = [dict[dict_keys[0]], dict[dict_keys[1]]]
    minx = min(temp)
    if minx == temp[0]:
        miny = dict_keys[0]
    else:
        miny = dict_keys[1]
    return [minx, miny]


# 转换成四点坐标
def rotatePoint(xc, yc, xp, yp, theta):
    xoff = xp - xc
    yoff = yp - yc
    cosTheta = math.cos(math.radians(theta))
    sinTheta = math.sin(math.radians(theta))
    pResx = cosTheta * xoff - sinTheta * yoff
    pResy = sinTheta * xoff + cosTheta * yoff
    # pRes = (xc + pResx, yc + pResy)
    # 保留一位小数点
    return float(format(xc + pResx, '.1f')), float(format(yc + pResy, '.1f'))


if __name__ == '__main__':
    root_path = 'D:/racedata\detrota/2_Aligned_RGB_infrared/anno-H'
    file_list = os.listdir(root_path)
    for i in range(0, len(file_list)):
        if ('.xml' in file_list[i]) or ('.XML' in file_list[i]):
            voc_to_dota(root_path, file_list[i])
            print('----------------------------------------{}{}----------------------------------------'
                  .format(file_list[i], ' has Done!'))

        else:
            print(file_list[i] + ' is not xml file')
    print(all_name)

训练

训练配置文件修改，一个是你的类别数目和数据的位置，和数据类别

dataset_type = 'DOTADataset'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

init_cls = ['car','bus', 'light-truck', 'heavy-truck']
data = dict(
    samples_per_gpu=8,
    workers_per_gpu=8,
    train=dict(
        type='DOTADataset',
        ann_file='data2/anno-H-txt-new/',
        classes=['car','bus', 'light-truck', 'heavy-truck'],
        img_prefix='data2/train/image-H/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='LoadAnnotations', with_bbox=True),
            dict(type='RResize', img_scale=(640, 640)),
            dict(
                type='RRandomFlip',
                flip_ratio=[0.25, 0.25, 0.25],
                direction=['horizontal', 'vertical', 'diagonal'],
                version='oc'),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_rgb=True),
            dict(type='Pad', size_divisor=32),
            dict(type='DefaultFormatBundle'),
            dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
        ],
        version='oc'),
    val=dict(
        type='DOTADataset',
        ann_file='data2/anno-H-txt-new/',
        classes=['car','bus', 'light-truck', 'heavy-truck'],
        img_prefix='data2/train/image-H/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(
                type='MultiScaleFlipAug',
                img_scale=(640, 640),
                flip=False,
                transforms=[
                    dict(type='RResize'),
                    dict(
                        type='Normalize',
                        mean=[123.675, 116.28, 103.53],
                        std=[58.395, 57.12, 57.375],
                        to_rgb=True),
                    dict(type='Pad', size_divisor=32),
                    dict(type='DefaultFormatBundle'),
                    dict(type='Collect', keys=['img'])
                ])
        ],
        version='oc'),
    test=dict(
        type='DOTADataset',
        ann_file='data2/anno-H-txt-new/',
        classes=['car','bus', 'light-truck', 'heavy-truck'],
        img_prefix='data2/train/image-H/',
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(
                type='MultiScaleFlipAug',
                img_scale=(640, 640),
                flip=False,
                transforms=[
                    dict(type='RResize'),
                    dict(
                        type='Normalize',
                        mean=[123.675, 116.28, 103.53],
                        std=[58.395, 57.12, 57.375],
                        to_rgb=True),
                    dict(type='Pad', size_divisor=32),
                    dict(type='DefaultFormatBundle'),
                    dict(type='Collect', keys=['img'])
                ])
        ],
        version='oc'))
evaluation = dict(interval=12, metric='mAP')
optimizer = dict(
    #_delete_=True,
    type='AdamW',
    lr=0.0001,
    betas=(0.9, 0.999),
    weight_decay=0.05,
    paramwise_cfg=dict(
        custom_keys={
            'absolute_pos_embed': dict(decay_mult=0.),
            'relative_position_bias_table': dict(decay_mult=0.),
            'norm': dict(decay_mult=0.)
        }))
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.3333333333333333,
    step=[15, 19])
runner = dict(type='EpochBasedRunner', max_epochs=24)
checkpoint_config = dict(interval=12)
log_config = dict(interval=50, hooks=[dict(type='TextLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
#load_from = '/home/a1804/cp/mmrotate/weights/rotated_retinanet_obb_r50_fpn_fp16_1x_dota_le90-01de71b5.pth'
resume_from = None
workflow = [('train', 1)]
opencv_num_threads = 0
mp_start_method = 'fork'
angle_version = 'oc'
custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
checkpoint_file = 'https://download.openmmlab.com/mmclassification/v0/convnext/downstream/convnext-tiny_3rdparty_32xb128-noema_in1k_20220301-795e9634.pth'

model = dict(
    type='RotatedRetinaNet',
    backbone=dict(
        #_delete_=True,
        type='mmcls.ConvNeXt',
        arch='tiny',
        out_indices=[0, 1, 2, 3],
        drop_path_rate=0.4,
        layer_scale_init_value=1.0,
        gap_before_final_norm=False,
        init_cfg=dict(
            type='Pretrained', checkpoint=checkpoint_file,
            prefix='backbone.')),
    neck=dict(
        type='FPN',
        in_channels=[96, 192, 384, 768],
        out_channels=256,
        start_level=1,
        add_extra_convs='on_input',
        num_outs=5),
    bbox_head=dict(
        type='RotatedRetinaHead',
        num_classes=4,
        in_channels=256,
        stacked_convs=4,
        feat_channels=256,
        assign_by_circumhbbox=None,
        anchor_generator=dict(
            type='RotatedAnchorGenerator',
            octave_base_scale=4,
            scales_per_octave=3,
            ratios=[1.0, 0.5, 2.0],
            strides=[8, 16, 32, 64, 128]),
        bbox_coder=dict(
            type='DeltaXYWHAOBBoxCoder',
            angle_range='oc',
            norm_factor=None,
            edge_swap=False,
            proj_xy=False,
            target_means=(0.0, 0.0, 0.0, 0.0, 0.0),
            target_stds=(1.0, 1.0, 1.0, 1.0, 1.0)),
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
    train_cfg=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.4,
            min_pos_iou=0,
            ignore_iof_thr=-1,
            iou_calculator=dict(type='RBboxOverlaps2D')),
        allowed_border=-1,
        pos_weight=-1,
        debug=False),
    test_cfg=dict(
        nms_pre=2000,
        min_bbox_size=0,
        score_thr=0.05,
        nms=dict(iou_thr=0.1),
        max_per_img=2000))
work_dir = './new_dirs/baseline-convtiny'
auto_resume = False
gpu_ids = range(0, 1)

可视化

单张可视化

from argparse import ArgumentParser

from mmdet.apis import inference_detector, init_detector, show_result_pyplot

import mmrotate  # noqa: F401


config = 'new_dirs/baseline-convtiny/test.py'
checkpoint = 'new_dirs/baseline-convtiny/epoch_24.pth'
model = init_detector(config, checkpoint, device='cpu')
img='/home/a1804/cp/mmrotate/data2/val/Infrared/DJI_2022090101_0096.png'
result = inference_detector(model, img)

print(result)
print("++++++++++++++++++++")
print(result[0][:][4])
show_result_pyplot(
        model,
        img,
        result,
        score_thr=0.3,
        out_file='showdir/1.png'
        )

根据推理生成的json进行全部可视化预测

import cv2
import matplotlib.pyplot as plt
import numpy as np
import math
import json

def rotatePoint(xc, yc, xp, yp, theta):
    xoff = xp - xc
    yoff = yp - yc
    cosTheta = math.cos(math.radians(theta))
    sinTheta = math.sin(math.radians(theta))
    pResx = cosTheta * xoff - sinTheta * yoff
    pResy = sinTheta * xoff + cosTheta * yoff
    # pRes = (xc + pResx, yc + pResy)
    # 保留一位小数点
    return float(format(xc + pResx, '.1f')), float(format(yc + pResy, '.1f'))


with open('result.json','r') as f:
    data=json.load(f)
for img_data in data:
    reline=img_data['result']
    name=img_data['img_filename']
    if name[0]=="I":
        name=name[8:-4]+'.png'
    else:
        continue
    print(name)
    img=cv2.imread('data2/val/Infrared/'+name)
    all_data=[]
    for line in reline:
        cx,cy,w,h,angle=int(line['rbox'][0]),int(line['rbox'][1]),int(line['rbox'][2]),int(line['rbox'][3]),float(line['rbox'][4])
        x1 = int(cx - w / 2)
        y1 = int(cy - h / 2)
        x2 = int(cx + w / 2)
        y2 = int(cy + h / 2)
        # img = cv2.rectangle(img, (x1, y1), (x2, y2), (255, 0, 0), 2)
        # img = cv2.putText(img, str(int(angle)), (x1, y1), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0), 2)
        p0x, p0y = rotatePoint(cx, cy, cx - w / 2, cy - h / 2, -angle)
        p1x, p1y = rotatePoint(cx, cy, cx + w / 2, cy - h / 2, -angle)
        p2x, p2y = rotatePoint(cx, cy, cx + w / 2, cy + h / 2, -angle)
        p3x, p3y = rotatePoint(cx, cy, cx - w / 2, cy + h / 2, -angle)
        all_data.append(np.array([[int(p0x), int(p0y)], [int(p1x), int(p1y)], [int(p2x), int(p2y)], [int(p3x), int(p3y)]]))
    cv2.polylines(img, all_data, True, (0, 0, 255), 2)
    cv2.imwrite('showdir/' +name, img)
    

推理代码

这是一个比赛的提交格式，mmrotate预测生成的坐标信息是cx cy w h 弧度制预测 score，我们需要将弧度制预测*180/pi 算出具体的弧度就可以了。

from argparse import ArgumentParser
import json
from mmdet.apis import inference_detector, init_detector, show_result_pyplot
import math
import mmrotate  # noqa: F401
import os
config = 'new_dirs/baseline-convtiny/test.py'
checkpoint = 'new_dirs/baseline-convtiny/epoch_24.pth'
model = init_detector(config, checkpoint, device='cuda:2')
allimg=os.listdir('data2/val/Infrared/')
root='data2/val/Infrared/'
result_json=[]
#change_cate={0:0,1:2,2:3,3:2}
def make_json(result,img):
  le=len(result)
  temp_all={"img_filename":'0',"result":[]}
  temp_all2={"img_filename":'0',"result":[]}
  temp_all['img_filename']='RGB/'+img+'.jpg'
  temp_all2['img_filename']='Infrared/'+img+'.jpg'
  for i in range(le):
    for j in range(len(result[i])):
      temp_box={"category_id": 0,"rbox": [0,0,0,0,0],"score": 0}
      x,y,width,height,angle,score=result[i][j]
      if score<0.3:
        continue
      angle=angle*180/math.pi
      temp_box['category_id']=int(i)
      temp_box['rbox']=[float(x),float(y),float(width),float(height),360-float(angle)]
      temp_box['score']=float(score)
      temp_all['result'].append(temp_box)
      temp_all2['result'].append(temp_box)
  result_json.append(temp_all)
  result_json.append(temp_all2)
for img in allimg:
  print(img)
  path_img=root+img
  result = inference_detector(model,path_img)
  make_json(result,img[:-4])
with  open("./result.json", 'w')  as f:
    json.dump(result_json, f, indent=4)