文章目录
使用picodet进行小目标检测。
本文以检测小目标乒乓球为例，包括数据的准备及训练全流程。
1、数据准备
这次有两个分辩率的输入分别是416和256，对应的切图多尺度是这么做的：
416：[320, 352, 384, 416, 448, 480, 512]，对应configs/picodet/picodet_s_416_voc_npu_pingpang_20230322.yml这个配置文件
256: [160, 192, 224, 256, 288, 320, 352]，对应configs/picodet/picodet_s_416_voc_npu_pingpang_20230322.yml这个配置文件
以上各自的分辩率。
配置文件的多尺度如下：
batch_transforms:
# - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512], random_size: True, random_interp: True, keep_ratio: False}
- BatchRandomResize: {target_size: [160, 192, 224, 256, 288, 320, 352], random_size: True, random_interp: True, keep_ratio: False}
数据准备流程是：
voc->coco->sahi切图->转voc->生成训练数据
官方的主流是支持coco的，但我们使用voc是为了使用labelimg进行查看
原始数据位置：dataset/pqdetection_voc
pqdetection_voc
|----images
| |----a.jpg
| |----b.jpg
|----labels
| |----a.xml
| |----ab.xml
|----label_list.txt #写有类别名称
1.1 VOC转COCO
这个有我的其它博客链接，这里写一部分单独的代码
import os
import glob
import json
from tqdm import tqdm
import xml.etree.ElementTree as ET
def voc_xmls_to_cocojson(annpath,imgpath,labelpath, output_dir, output_file):
"""voc xmls 转cocojson,只生成一个json的文件
Args:
annpath (_type_): 标注文件xml的路径
imgpath (_type_): 图片的路径
labelpath (_type_): 标签文件的路径
output_dir (_type_): json保存路径
output_file (_type_): json文件名
"""
img_names = os.listdir(imgpath)
annotation_paths = glob.glob(os.path.join(annpath,'*.xml'))
with open(labelpath, 'r') as f:
labels_str = f.read().split()
labels_ids = list(range(1, len(labels_str) + 1))
label2id=dict(zip(labels_str, labels_ids))
output_json_dict = {
"images": [],
"type": "instances",
"annotations": [],
"categories": []
}
bnd_id = 1 # bounding box start id
im_id = 0
print('Start converting !')
for xml_path in tqdm(annotation_paths): # 在这里我们以xml为基础，而不是图片，因为有些图片没有目标，没有生成xml文件；这样必须用到glob，这会一定程度上减小程序运行速度
# Read annotation xml
im_dir = glob.glob(os.path.join(imgpath,os.path.splitext(os.path.basename(xml_path))[0]+".*"))[0]
img_name = os.path.basename(im_dir)
ann_tree = ET.parse(xml_path)
ann_root = ann_tree.getroot()
size = ann_root.find('size')
width = float(size.findtext('width'))
height = float(size.findtext('height'))
img_info = {
'file_name': img_name,
'height': height,
'width': width,
'id': im_id
}
# temp={"file_name":}
output_json_dict['images'].append(img_info)
for obj in ann_root.findall('object'):
label = obj.findtext('name')
assert label in label2id, "label is not in label2id."
category_id = label2id[label]
bndbox = obj.find('bndbox')
xmin = float(bndbox.findtext('xmin'))
ymin = float(bndbox.findtext('ymin'))
xmax = float(bndbox.findtext('xmax'))
ymax = float(bndbox.findtext('ymax'))
assert xmax > xmin and ymax > ymin, "Box size error."
o_width = xmax - xmin
o_height = ymax - ymin
ann = {
'area': o_width * o_height,
'iscrowd': 0,
'bbox': [xmin, ymin, o_width, o_height],
'category_id': category_id,
'ignore': 0,
}
ann.update({'image_id': im_id, 'id': bnd_id})
output_json_dict['annotations'].append(ann)
bnd_id = bnd_id + 1
im_id += 1
for label, label_id in label2id.items():
category_info = {'supercategory': 'none', 'id': label_id, 'name': label}
output_json_dict['categories'].append(category_info)
output_file = os.path.join(output_dir, output_file)
if not os.path.exists(output_dir):
os.makedirs(output_dir,exist_ok=True)
with open(output_file, 'w',encoding='utf-8') as f:
# output_json = json.dumps(output_json_dict)
# f.write(output_json)
json.dump(output_json_dict, f, indent=4, ensure_ascii=False) #支持中文
voc_xmls_to_cocojson(annpath='dataset/pqdetection_voc/labels/',
imgpath="dataset/pqdetection_voc/images",
labelpath="dataset/pqdetection_voc/label_list.txt",
output_dir="dataset/pqdetection_coco/",
output_file='pddetection.json')
Start converting !
100%|██████████| 1477/1477 [00:00<00:00, 11965.53it/s]
查看转换后数据的状态，sahi有许多个的关于coco的函数，可以查看github相关链接
from sahi.utils.coco import Coco
coco=Coco.from_coco_dict_or_path("dataset/pqdetection_coco/pddetection.json")
coco.stats
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 1477/1477 [00:00<00:00, 5365.99it/s]
{'num_images': 1477,
'num_annotations': 3924,
'num_categories': 1,
'num_negative_images': 0,
'num_images_per_category': {'ball': 1477},
'num_annotations_per_category': {'ball': 3924},
'min_num_annotations_in_image': 1,
'max_num_annotations_in_image': 25,
'avg_num_annotations_in_image': 2.6567366283006093,
'min_annotation_area': 36,
'max_annotation_area': 1158852,
'avg_annotation_area': 5425.130988786952,
'min_annotation_area_per_category': {'ball': 36},
'max_annotation_area_per_category': {'ball': 1158852}}
2、使用sahi切图
切图有两个尺度的，一个是416，一个是256
from sahi.slicing import slice_coco
from sahi.utils.file import load_json
from PIL import Image,ImageDraw
import matplotlib.pyplot as plt
2.1 切图分析及过程可视化
尝试其中的一张图像
coco_dict = load_json('dataset/pqdetection_coco/pddetection.json')
f,axarr = plt.subplots(1,1,figsize=(12,12))
#read img
img_ind=100
img = Image.open(os.path.join('dataset/pqdetection_voc/images',coco_dict['images'][img_ind]["file_name"]))
# iter annotations
for ann_ind in range(len(coco_dict['annotations'])):
if coco_dict["annotations"][ann_ind]['image_id']==coco_dict['images'][img_ind]['id']:
xywh = coco_dict["annotations"][ann_ind]["bbox"]
xyxy=[xywh[0],xywh[1],xywh[0]+xywh[2],xywh[1]+xywh[3]]
ImageDraw.Draw(img).rectangle(xyxy,width=2)
axarr.imshow(img)
上图标注出了乒乓球位置，是个白框;更多的切图效果可以查看https://github.com/obss/sahi/blob/main/demo/slicing.ipynb
2.2 使用完整的切图命令进行切图
进行切图，图片大小256x256,重叠0.2
coco_dict,coco_path = slice_coco(
coco_annotation_file_path="dataset/pqdetection_coco/pddetection.json",
image_dir='dataset/pqdetection_voc/images',
output_coco_annotation_file_name='sliced_coco.json',
ignore_negative_samples=False,
output_dir='sliced_coco',
slice_height=256,
slice_width=256,
overlap_height_ratio=0.2,
overlap_width_ratio=0.2,
min_area_ratio=0.1,
verbose=True)
以上代码如果图片多,最好放到一个脚本里执行。其中min_arae_ratio的定义是(标注框面积/crop图面积)/(标注框面积/原图面积)<min_area_ration那么标注就相当与没有。
ignore_negative_samples这个参数是指切后的图如果图片没有标注，是否进行忽略，默认是False,即不忽略，我们如果只关注有标注框的图那么可以设置成True.
以上切图工作会花一些时间。
文件夹保存结果为
sliced_coco:
2023_0302_164336_005_510_0_0_256_256.jpg
2023_0302_164336_005_510_0_205_256_461.jpg等
后边是切图的时候左上右下的结果
还有一个
sliced_coco.json_coco.json的文件，看来是我们指定的output_coco_annotation_file_name=‘sliced_coco.json’,后边又加一个"_coco.json"的结果，
所以我们定义名字的时候可以不这么做。
查看一上切图后的数据集状态：
from sahi.utils.coco import Coco
coco=Coco.from_coco_dict_or_path("sliced_coco/sliced_coco.json_coco.json")
coco.stats
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 34451/34451 [00:01<00:00, 33605.44it/s]
{'num_images': 34451,
'num_annotations': 10429,
'num_categories': 1,
'num_negative_images': 27773,
'num_images_per_category': {'ball': 6678},
'num_annotations_per_category': {'ball': 10429},
'min_num_annotations_in_image': 0,
'max_num_annotations_in_image': 11,
'avg_num_annotations_in_image': 1.5616951182988918,
'min_annotation_area': 9,
'max_annotation_area': 65536,
'avg_annotation_area': 3454.7506951769105,
'min_annotation_area_per_category': {'ball': 9},
'max_annotation_area_per_category': {'ball': 65536}}
将ignore_negative_samples设置成True
coco_dict,coco_path = slice_coco(
coco_annotation_file_path="dataset/pqdetection_coco/pddetection.json",
image_dir='dataset/pqdetection_voc/images',
output_coco_annotation_file_name='sliced1',
ignore_negative_samples=False,
output_dir='sliced1_coco',
slice_height=256,
slice_width=256,
overlap_height_ratio=0.2,
overlap_width_ratio=0.2,
min_area_ratio=0.1,
verbose=True)
接着查看状态：
from sahi.utils.coco import Coco
coco=Coco.from_coco_dict_or_path("sliced_coco1/sliced1_coco.json")
coco.stats
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 6678/6678 [00:00<00:00, 8963.77it/s]
{'num_images': 6678,
'num_annotations': 10429,
'num_categories': 1,
'num_negative_images': 0,
'num_images_per_category': {'ball': 6678},
'num_annotations_per_category': {'ball': 10429},
'min_num_annotations_in_image': 1,
'max_num_annotations_in_image': 11,
'avg_num_annotations_in_image': 1.5616951182988918,
'min_annotation_area': 9,
'max_annotation_area': 65536,
'avg_annotation_area': 3454.7506951769105,
'min_annotation_area_per_category': {'ball': 9},
'max_annotation_area_per_category': {'ball': 65536}}
paddle 自带命令"python tools/slice_image.py --image_dir --json_path --output_dir --slice_size --overlap_ratio"，但少了ignore_negative_samples,这个参数对与密集小目标可以设置成False,但我们的小目标图在原始图上只有一小部分，所在背景图太多，设置成True更合适，关于slice命令的参数，如下：
from sahi.scripts.slice_coco import slice
help(slice)
Help on function slice in module sahi.scripts.slice_coco:
slice(image_dir: str, dataset_json_path: str, slice_size: int = 512, overlap_ratio: float = 0.2, ignore_negative_samples: bool = False, output_dir: str = 'runs/slice_coco', min_area_ratio: float = 0.1)
Args:
image_dir (str): directory for coco images
dataset_json_path (str): file path for the coco dataset json file
slice_size (int)
overlap_ratio (float): slice overlap ratio
ignore_negative_samples (bool): ignore images without annotation
output_dir (str): output export dir
min_area_ratio (float): If the cropped annotation area to original
annotation ratio is smaller than this value, the annotation
is filtered out. Default 0.1.
切图命令如下，其中overlap_ratio越大，分割的越密集也就会越慢，数据集特别小，标注框附近背景变化大的时候可以把这个参数设置大。这个命令会只显示进度条，用来查看进度
slice(
image_dir="dataset/pqdetection_voc/images/",
dataset_json_path="dataset/pqdetection_coco/pddetection.json",
output_dir="dataset/pqdetection_sliced/pq_256_075",
slice_size=256,
overlap_ratio=0.75,
ignore_negative_samples=True,
min_area_ratio=0.1 )
生成文件结构是
dataset/pqdetection_sliced/pq_256_075
|
|----pddetections_images_256_075/*.jpg
|
|----pddetection_256_075.json
文件的状态是：
from sahi.utils.coco import Coco
coco=Coco.from_coco_dict_or_path("dataset/pqdetection_sliced/pq_256_075/pddetection_256_075.json")
coco.stats
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 44695/44695 [00:05<00:00, 8403.17it/s]
{'num_images': 44695,
'num_annotations': 67009,
'num_categories': 1,
'num_negative_images': 0,
'num_images_per_category': {'ball': 44695},
'num_annotations_per_category': {'ball': 67009},
'min_num_annotations_in_image': 1,
'max_num_annotations_in_image': 12,
'avg_num_annotations_in_image': 1.4992504754446807,
'min_annotation_area': 6,
'max_annotation_area': 65536,
'avg_annotation_area': 3860.9125938306793,
'min_annotation_area_per_category': {'ball': 6},
'max_annotation_area_per_category': {'ball': 65536}}
同理完成其它尺度的切割，过程比较慢，可以用个多进程来完成 以下是我最终使用的代码
from multiprocessing import Process
import os
sizes=[160,192,224,256,288,320,352,384,416,448,480,512]
ratios=[0.75]*len(sizes)
out_size_ratio=[[f"dataset/pqdetection_sliced/pq_{size}_{int(ratio*100)}",size,ratio] for size,ratio in zip(sizes,ratios)]
def slice_my(image_dir="dataset/pqdetection_voc/images/",
dataset_json_path="dataset/pqdetection_coco/pddetection.json",
output_dir="dataset/pqdetection_sliced/pq_256_075",
slice_size=256,
overlap_ratio=0.75,
ignore_negative_samples=True,
min_area_ratio=0.1 ):
print(f"output_dir:{output_dir} pid:{os.getpid()}") #只是想加一个进程id的获取，方便把进程手动停止
slice(
image_dir=image_dir,
dataset_json_path=dataset_json_path,
output_dir=output_dir,
slice_size=slice_size,
overlap_ratio=overlap_ratio,
ignore_negative_samples=ignore_negative_samples,
min_area_ratio= min_area_ratio)
out_size_ratio
[['dataset/pqdetection_sliced/pq_160_75', 160, 0.75],
['dataset/pqdetection_sliced/pq_192_75', 192, 0.75],
['dataset/pqdetection_sliced/pq_224_75', 224, 0.75],
['dataset/pqdetection_sliced/pq_256_75', 256, 0.75],
['dataset/pqdetection_sliced/pq_288_75', 288, 0.75],
['dataset/pqdetection_sliced/pq_320_75', 320, 0.75],
['dataset/pqdetection_sliced/pq_352_75', 352, 0.75],
['dataset/pqdetection_sliced/pq_384_75', 384, 0.75],
['dataset/pqdetection_sliced/pq_416_75', 416, 0.75],
['dataset/pqdetection_sliced/pq_448_75', 448, 0.75],
['dataset/pqdetection_sliced/pq_480_75', 480, 0.75],
['dataset/pqdetection_sliced/pq_512_75', 512, 0.75]]
kwargs=dict(
image_dir="dataset/pqdetection_voc/images/",
dataset_json_path="dataset/pqdetection_coco/pddetection.json",
output_dir=out_size_ratio[0][0],
slice_size=out_size_ratio[0][1],
overlap_ratio=out_size_ratio[0][2],
ignore_negative_samples=True,
min_area_ratio=0.1 )
kwargs
{'image_dir': 'dataset/pqdetection_voc/images/',
'dataset_json_path': 'dataset/pqdetection_coco/pddetection.json',
'output_dir': 'dataset/pqdetection_sliced/pq_160_75',
'slice_size': 160,
'overlap_ratio': 0.75,
'ignore_negative_samples': True,
'min_area_ratio': 0.1}
processer=[]
for i in range(len(out_size_ratio)):
kwargs=dict(
image_dir="dataset/pqdetection_voc/images/",
dataset_json_path="dataset/pqdetection_coco/pddetection.json",
output_dir=out_size_ratio[i][0],
slice_size=out_size_ratio[i][1],
overlap_ratio=out_size_ratio[i][2],
ignore_negative_samples=True,
min_area_ratio=0.1 )
p = Process(target=slice_my,kwargs=kwargs)
processer.append(p)
p.start()
for p in processer: #等待所有进程都结速
p.join()
以上输出会比较乱,在jupyter中执行会有问题，要放到python脚本中执行.最后生成的文件结构是：
pqdetection_sliced
├── pq_160_75
│ ├── pddetection_160_075.json
│ └── pddetection_images_160_075
├── pq_192_75
│ ├── pddetection_192_075.json
│ └── pddetection_images_192_075
├── pq_224_75
│ ├── pddetection_224_075.json
│ └── pddetection_images_224_075
├── pq_256_75
│ ├── pddetection_256_075.json
│ └── pddetection_images_256_075
├── pq_288_75
│ ├── pddetection_288_075.json
│ └── pddetection_images_288_075
├── pq_320_75
│ ├── pddetection_320_075.json
│ └── pddetection_images_320_075
├── pq_352_75
│ ├── pddetection_352_075.json
│ └── pddetection_images_352_075
├── pq_384_75
│ ├── pddetection_384_075.json
│ └── pddetection_images_384_075
├── pq_416_75
│ ├── pddetection_416_075.json
│ └── pddetection_images_416_075
├── pq_448_75
│ ├── pddetection_448_075.json
│ └── pddetection_images_448_075
├── pq_480_75
│ ├── pddetection_480_075.json
│ └── pddetection_images_480_075
└── pq_512_75
├── pddetection_512_075.json
└── pddetection_images_512_075
25 directories, 12 files
2.3 对各个数据集的状态进行查看
import os
import glob
from pprint import pprint
from sahi.utils.coco import Coco
for i in out_size_ratio:
jsonpath = glob.glob(os.path.join(i[0],'*.json'))[0]
coco=Coco.from_coco_dict_or_path(jsonpath)
pprint(f"{os.path.basename(i[0])}:")
pprint(coco.stats)
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 60068/60068 [00:05<00:00, 10029.56it/s]
'pq_160_75:'
{'avg_annotation_area': 2752.9663472653942,
'avg_num_annotations_in_image': 1.296097755876673,
'max_annotation_area': 25600,
'max_annotation_area_per_category': {'ball': 25600},
'max_num_annotations_in_image': 9,
'min_annotation_area': 6,
'min_annotation_area_per_category': {'ball': 6},
'min_num_annotations_in_image': 1,
'num_annotations': 77854,
'num_annotations_per_category': {'ball': 77854},
'num_categories': 1,
'num_images': 60068,
'num_images_per_category': {'ball': 60068},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 52959/52959 [00:05<00:00, 9287.43it/s]
'pq_192_75:'
{'avg_annotation_area': 3227.159324899192,
'avg_num_annotations_in_image': 1.3626956702354651,
'max_annotation_area': 36864,
'max_annotation_area_per_category': {'ball': 36864},
'max_num_annotations_in_image': 10,
'min_annotation_area': 7,
'min_annotation_area_per_category': {'ball': 7},
'min_num_annotations_in_image': 1,
'num_annotations': 72167,
'num_annotations_per_category': {'ball': 72167},
'num_categories': 1,
'num_images': 52959,
'num_images_per_category': {'ball': 52959},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 47829/47829 [00:05<00:00, 8223.08it/s]
'pq_224_75:'
{'avg_annotation_area': 3784.9258265381764,
'avg_num_annotations_in_image': 1.41023228585168,
'max_annotation_area': 50176,
'max_annotation_area_per_category': {'ball': 50176},
'max_num_annotations_in_image': 10,
'min_annotation_area': 6,
'min_annotation_area_per_category': {'ball': 6},
'min_num_annotations_in_image': 1,
'num_annotations': 67450,
'num_annotations_per_category': {'ball': 67450},
'num_categories': 1,
'num_images': 47829,
'num_images_per_category': {'ball': 47829},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 44695/44695 [00:05<00:00, 8464.05it/s]
'pq_256_75:'
{'avg_annotation_area': 3860.9125938306793,
'avg_num_annotations_in_image': 1.4992504754446807,
'max_annotation_area': 65536,
'max_annotation_area_per_category': {'ball': 65536},
'max_num_annotations_in_image': 12,
'min_annotation_area': 6,
'min_annotation_area_per_category': {'ball': 6},
'min_num_annotations_in_image': 1,
'num_annotations': 67009,
'num_annotations_per_category': {'ball': 67009},
'num_categories': 1,
'num_images': 44695,
'num_images_per_category': {'ball': 44695},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 38428/38428 [00:04<00:00, 8022.50it/s]
'pq_288_75:'
{'avg_annotation_area': 4340.086797086578,
'avg_num_annotations_in_image': 1.5470230040595399,
'max_annotation_area': 82944,
'max_annotation_area_per_category': {'ball': 82944},
'max_num_annotations_in_image': 13,
'min_annotation_area': 7,
'min_annotation_area_per_category': {'ball': 7},
'min_num_annotations_in_image': 1,
'num_annotations': 59449,
'num_annotations_per_category': {'ball': 59449},
'num_categories': 1,
'num_images': 38428,
'num_images_per_category': {'ball': 38428},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 33576/33576 [00:04<00:00, 7668.71it/s]
'pq_320_75:'
{'avg_annotation_area': 4339.065280923504,
'avg_num_annotations_in_image': 1.6305694543721705,
'max_annotation_area': 102400,
'max_annotation_area_per_category': {'ball': 102400},
'max_num_annotations_in_image': 14,
'min_annotation_area': 6,
'min_annotation_area_per_category': {'ball': 6},
'min_num_annotations_in_image': 1,
'num_annotations': 54748,
'num_annotations_per_category': {'ball': 54748},
'num_categories': 1,
'num_images': 33576,
'num_images_per_category': {'ball': 33576},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 34133/34133 [00:04<00:00, 7323.31it/s]
'pq_352_75:'
{'avg_annotation_area': 4425.495999587941,
'avg_num_annotations_in_image': 1.7063838514047989,
'max_annotation_area': 123904,
'max_annotation_area_per_category': {'ball': 123904},
'max_num_annotations_in_image': 16,
'min_annotation_area': 8,
'min_annotation_area_per_category': {'ball': 8},
'min_num_annotations_in_image': 1,
'num_annotations': 58244,
'num_annotations_per_category': {'ball': 58244},
'num_categories': 1,
'num_images': 34133,
'num_images_per_category': {'ball': 34133},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 28585/28585 [00:04<00:00, 6942.04it/s]
'pq_384_75:'
{'avg_annotation_area': 4594.21496125687,
'avg_num_annotations_in_image': 1.801399335315725,
'max_annotation_area': 147456,
'max_annotation_area_per_category': {'ball': 147456},
'max_num_annotations_in_image': 17,
'min_annotation_area': 8,
'min_annotation_area_per_category': {'ball': 8},
'min_num_annotations_in_image': 1,
'num_annotations': 51493,
'num_annotations_per_category': {'ball': 51493},
'num_categories': 1,
'num_images': 28585,
'num_images_per_category': {'ball': 28585},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 24423/24423 [00:03<00:00, 7061.30it/s]
'pq_416_75:'
{'avg_annotation_area': 5258.0641016721165,
'avg_num_annotations_in_image': 1.7655079228595996,
'max_annotation_area': 173056,
'max_annotation_area_per_category': {'ball': 173056},
'max_num_annotations_in_image': 16,
'min_annotation_area': 9,
'min_annotation_area_per_category': {'ball': 9},
'min_num_annotations_in_image': 1,
'num_annotations': 43119,
'num_annotations_per_category': {'ball': 43119},
'num_categories': 1,
'num_images': 24423,
'num_images_per_category': {'ball': 24423},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 24265/24265 [00:03<00:00, 6604.52it/s]
'pq_448_75:'
{'avg_annotation_area': 4893.795871259963,
'avg_num_annotations_in_image': 1.8925200906655677,
'max_annotation_area': 200704,
'max_annotation_area_per_category': {'ball': 200704},
'max_num_annotations_in_image': 16,
'min_annotation_area': 13,
'min_annotation_area_per_category': {'ball': 13},
'min_num_annotations_in_image': 1,
'num_annotations': 45922,
'num_annotations_per_category': {'ball': 45922},
'num_categories': 1,
'num_images': 24265,
'num_images_per_category': {'ball': 24265},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 18265/18265 [00:02<00:00, 7345.95it/s]
'pq_480_75:'
{'avg_annotation_area': 5578.763276481661,
'avg_num_annotations_in_image': 1.9196277032575966,
'max_annotation_area': 230400,
'max_annotation_area_per_category': {'ball': 230400},
'max_num_annotations_in_image': 17,
'min_annotation_area': 8,
'min_annotation_area_per_category': {'ball': 8},
'min_num_annotations_in_image': 1,
'num_annotations': 35062,
'num_annotations_per_category': {'ball': 35062},
'num_categories': 1,
'num_images': 18265,
'num_images_per_category': {'ball': 18265},
'num_negative_images': 0}
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 17726/17726 [00:02<00:00, 6080.24it/s]
'pq_512_75:'
{'avg_annotation_area': 5312.11180071894,
'avg_num_annotations_in_image': 2.0088006318402347,
'max_annotation_area': 262144,
'max_annotation_area_per_category': {'ball': 262144},
'max_num_annotations_in_image': 17,
'min_annotation_area': 7,
'min_annotation_area_per_category': {'ball': 7},
'min_num_annotations_in_image': 1,
'num_annotations': 35608,
'num_annotations_per_category': {'ball': 35608},
'num_categories': 1,
'num_images': 17726,
'num_images_per_category': {'ball': 17726},
'num_negative_images': 0}
2.4 过滤数据集中不合适的框
切图过程中必然会产生一些只切了一部分目标物体的图，这部分图要过滤掉
过滤主要：超过图片高宽、面积
import os
import glob
from sahi.utils.coco import Coco
from sahi.utils.coco import remove_invalid_coco_results
from sahi.utils.file import save_json
for i in out_size_ratio:
datapath = i[0]
print(f"Start do {datapath}")
jsonfile = glob.glob(os.path.join(datapath,'*.json'))[0]
jsonname = os.path.splitext(os.path.basename(jsonfile))[0]
savejson=os.path.join(datapath,jsonname+'_filtered.json')
coco = Coco.from_coco_dict_or_path(jsonfile)
print(f" init coco img_num:{len(coco.json['images'])} annottations num:{len(coco.json['annotations'])}")
#因为是标注的图片，所以用图片高宽对标注框直接进行clip
coco = coco.get_coco_with_clipped_bboxes()
#面积
intervals_per_category={"ball":{"min":9,"max":2500}}
area_filtered_coco=coco.get_area_filtered_coco(intervals_per_category=intervals_per_category)
print(f" area filtered coco img_num:{len(area_filtered_coco.json['images'])} annottations num:{len(area_filtered_coco.json['annotations'])}")
save_json(area_filtered_coco.json,savejson)
Start do dataset/pqdetection_sliced/pq_160_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 60068/60068 [00:06<00:00, 9621.37it/s]
init coco img_num:60068 annottations num:77854
area filtered coco img_num:47588 annottations num:62160
Start do dataset/pqdetection_sliced/pq_192_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 52959/52959 [00:06<00:00, 8777.43it/s]
init coco img_num:52959 annottations num:72167
area filtered coco img_num:42357 annottations num:58173
Start do dataset/pqdetection_sliced/pq_224_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 47829/47829 [00:05<00:00, 8405.05it/s]
init coco img_num:47829 annottations num:67450
area filtered coco img_num:38547 annottations num:54666
Start do dataset/pqdetection_sliced/pq_256_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 44695/44695 [00:05<00:00, 7904.13it/s]
init coco img_num:44695 annottations num:67009
area filtered coco img_num:36976 annottations num:55907
Start do dataset/pqdetection_sliced/pq_288_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 38428/38428 [00:05<00:00, 7582.49it/s]
init coco img_num:38428 annottations num:59449
area filtered coco img_num:31925 annottations num:49597
Start do dataset/pqdetection_sliced/pq_320_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 33576/33576 [00:04<00:00, 7037.32it/s]
init coco img_num:33576 annottations num:54748
area filtered coco img_num:28577 annottations num:46788
Start do dataset/pqdetection_sliced/pq_352_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 34133/34133 [00:04<00:00, 6937.45it/s]
init coco img_num:34133 annottations num:58244
area filtered coco img_num:29367 annottations num:50455
Start do dataset/pqdetection_sliced/pq_384_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 28585/28585 [00:04<00:00, 6415.28it/s]
init coco img_num:28585 annottations num:51493
area filtered coco img_num:24956 annottations num:45075
Start do dataset/pqdetection_sliced/pq_416_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 24423/24423 [00:03<00:00, 6390.16it/s]
init coco img_num:24423 annottations num:43119
area filtered coco img_num:21233 annottations num:37535
Start do dataset/pqdetection_sliced/pq_448_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 24265/24265 [00:03<00:00, 7433.75it/s]
init coco img_num:24265 annottations num:45922
area filtered coco img_num:21508 annottations num:40713
Start do dataset/pqdetection_sliced/pq_480_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 18265/18265 [00:03<00:00, 5657.12it/s]
init coco img_num:18265 annottations num:35062
area filtered coco img_num:16187 annottations num:30991
Start do dataset/pqdetection_sliced/pq_512_75
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 17726/17726 [00:03<00:00, 5450.05it/s]
init coco img_num:17726 annottations num:35608
area filtered coco img_num:15916 annottations num:31855
3、转换成VOC
转换voc一个是方便我们用labelimg查看标注的情况，另一个方便组合数据，当然也会增加存储空间。当然coco本身也可以进行组合，如下：
from sahi.utils.coco import Coco
from sahi.utils.file import save_json
coco_1 = Coco.from_coco_dict_or_path("dataset/pqdetection_sliced/pq_160_75/pddetection_160_075.json","dataset/pqdetection_sliced/pq_160_75/pddetection_images_160_075")
coco_2 = Coco.from_coco_dict_or_path("dataset/pqdetection_sliced/pq_192_75/pddetection_192_075.json","dataset/pqdetection_sliced/pq_192_75/pddetection_images_192_075")
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 60068/60068 [00:06<00:00, 8639.41it/s]
indexing coco dataset annotations...
Loading coco annotations: 100%|██████████| 52959/52959 [00:06<00:00, 8750.60it/s]
coco_1.merge(coco_2)
save_json(coco_1.json,"merged_coco.json")
'desired_name2id' is not specified, combining all categories.
Categories are formed as:
[{'id': 0, 'name': 'ball', 'supercategory': 'ball'}]
要注意的是原先的json中保存的文件名都是图片文件名本身，而合并后文件名变成了绝对路径,这个应该是防止存在同名文件的情况。如果合并json，最后再转换voc，如果有同名的也是不好处理,所以我们各自单独转换,这样即使文件名相同也不怕。最后图片的文件名是原文件名_xmin_ymin_xmax_ymax.jpg
from pycocotools.coco import COCO
import os, cv2, shutil
from lxml import etree, objectify
from tqdm import tqdm
from PIL import Image
import numpy as np
import time
import json
import argparse
def cover_copy(src,dst):
'''
src和dst都必须是文件，该函数是执行覆盖操作
'''
if os.path.exists(dst):
os.remove(dst)
shutil.copy(src,dst)
else:
shutil.copy(src,dst)
def coco2voc(imgpath='VOCdevkit/COCO_VOC',jsonpath=None,savepath='COCO'):
"""
savepath:用来存放转换后数据和标注文件
imgpath:用来指定原始COCO数据集的存放位置
"""
img_savepath= os.path.join(savepath,'images')
ann_savepath=os.path.join(savepath,'labels')
for p in [img_savepath,ann_savepath]:
if os.path.exists(p):
shutil.rmtree(p)
os.makedirs(p)
else:
os.makedirs(p)
start = time.time()
no_ann=[] #用来存放没有标注数据的图片的id,并将这些图片复制到results文件夹中
not_rgb=[] #是灰度图，同样将其保存
print('loading annotations into memory...')
tic = time.time()
with open(jsonpath, 'r') as f:
dataset_ann = json.load(f)
assert type(
dataset_ann
) == dict, 'annotation file format {} not supported'.format(
type(dataset))
print('Done (t={:0.2f}s)'.format(time.time() - tic))
coco = COCO(jsonpath)
classes = dict()
for cat in coco.dataset['categories']:
classes[cat['id']] = cat['name']
imgIds = coco.getImgIds()
# imgIds=imgIds[0:1000]#测试用，抽取10张图片，看下存储效果
for imgId in tqdm(imgIds):
img = coco.loadImgs(imgId)[0]
filename = img['file_name']
filepath=os.path.join(imgpath,filename)
annIds = coco.getAnnIds(imgIds=img['id'], iscrowd=None)
anns = coco.loadAnns(annIds)
if not len(anns):
# print(f"{dataset}:{imgId}该文件没有标注信息，将其复制到{dataset}_noann_result中，以使查看")
no_ann.append(imgId)
result_path = os.path.join(savepath,"noann_result")
dest_path = os.path.join(result_path,filename)
if not os.path.exists(result_path):
os.makedirs(result_path)
cover_copy(filepath,dest_path)
continue #如果没有标注信息，则把没有标注信息的图片移动到相关结果文件 noann_result中,来进行查看 ，然后返回做下一张图
#有标注信息，接着往下走，获取标注信息
objs = []
for ann in anns:
name = classes[ann['category_id']]
if 'bbox' in ann:
# print('bbox in ann',imgId)
bbox = ann['bbox']
xmin = (int)(bbox[0])
ymin = (int)(bbox[1])
xmax = (int)(bbox[2] + bbox[0])
ymax = (int)(bbox[3] + bbox[1])
obj = [name, 1.0, xmin, ymin, xmax, ymax]
#标错框在这里
if not(xmin-xmax==0 or ymin-ymax==0):
objs.append(obj)
else:
print(f"{imgId} bbox在标注文件中不存在")# 单张图有多个标注框，某个类别没有框
annopath = os.path.join(ann_savepath,filename[:-3] + "xml") #生成的xml文件保存路径
dst_path = os.path.join(img_savepath,filename)
im = Image.open(filepath)
image = np.array(im).astype(np.uint8)
if im.mode != "RGB":
# if img.shape[-1] != 3:
# print(f"{dataset}:{imgId}该文件非rgb图，其复制到{dataset}_notrgb_result中，以使查看")
# print(f"img.shape{image.shape} and img.mode{im.mode}")
not_rgb.append(imgId)
result_path = os.path.join(savepath,"notrgb_result")
dest_path = os.path.join(result_path,filename)
if not os.path.exists(result_path):
os.makedirs(result_path)
cover_copy(filepath,dest_path) #复制到notrgb_result来方便查看
im=im.convert('RGB')
image = np.array(im).astype(np.uint8)
im.save(dst_path,quality=95)#图片经过转换后，放到我们需要的位置片
im.close()
else:
cover_copy(filepath, dst_path)#把原始图像复制到目标文件夹
E = objectify.ElementMaker(annotate=False)
anno_tree = E.annotation(
E.folder('VOC'),
E.filename(filename),
E.source(
E.database('COCO'),
E.annotation('VOC'),
E.image('COCO')
),
E.size(
E.width(image.shape[1]),
E.height(image.shape[0]),
E.depth(image.shape[2])
),
E.segmented(0)
)
for obj in objs:
E2 = objectify.ElementMaker(annotate=False)
anno_tree2 = E2.object(
E.name(obj[0]),
E.pose(),
E.truncated("0"),
E.difficult(0),
E.bndbox(
E.xmin(obj[2]),
E.ymin(obj[3]),
E.xmax(obj[4]),
E.ymax(obj[5])
)
)
anno_tree.append(anno_tree2)
etree.ElementTree(anno_tree).write(annopath, pretty_print=True)
print(f"该数据集有{len(no_ann)}/{len(imgIds)}张图片没有instance标注信息，已经这些图片复制到{savepath}/noann_result中以使进行查看")
print(f"该数据集有{len(not_rgb)}/{len(imgIds)}张图片是非RGB图像，已经这些图片复制到{savepath}/notrgb_result中以使进行查看")
duriation = time.time()-start
print(f"数据集处理完成用时{round(duriation/60,2)}分")
for i in out_size_ratio:
coco_path = i[0]
name = os.path.basename(coco_path)
for p in os.listdir(coco_path):
path = os.path.join(coco_path,p)
if os.path.isfile(path) and path.endswith('filtered.json'):
jsonpath=path
elif os.path.isdir(path):
imgpath=path
coco2voc(imgpath,jsonpath,savepath='dataset/pqdetection_sliced2voc/'+name)
loading annotations into memory...
Done (t=0.20s)
loading annotations into memory...
Done (t=0.88s)
creating index...
index created!
100%|██████████| 47588/47588 [01:57<00:00, 404.14it/s]
该数据集有0/47588张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_160_75/noann_result中以使进行查看
该数据集有0/47588张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_160_75/notrgb_result中以使进行查看
数据集处理完成用时1.98分
loading annotations into memory...
Done (t=0.17s)
loading annotations into memory...
Done (t=0.73s)
creating index...
index created!
100%|██████████| 42357/42357 [02:11<00:00, 321.90it/s]
该数据集有0/42357张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_192_75/noann_result中以使进行查看
该数据集有0/42357张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_192_75/notrgb_result中以使进行查看
数据集处理完成用时2.21分
loading annotations into memory...
Done (t=0.16s)
loading annotations into memory...
Done (t=0.72s)
creating index...
index created!
100%|██████████| 38547/38547 [02:00<00:00, 320.49it/s]
该数据集有0/38547张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_224_75/noann_result中以使进行查看
该数据集有0/38547张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_224_75/notrgb_result中以使进行查看
数据集处理完成用时2.02分
loading annotations into memory...
Done (t=0.16s)
loading annotations into memory...
Done (t=0.75s)
creating index...
index created!
100%|██████████| 36976/36976 [02:43<00:00, 226.09it/s]
该数据集有0/36976张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_256_75/noann_result中以使进行查看
该数据集有0/36976张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_256_75/notrgb_result中以使进行查看
数据集处理完成用时2.74分
loading annotations into memory...
Done (t=0.14s)
loading annotations into memory...
Done (t=0.15s)
creating index...
index created!
100%|██████████| 31925/31925 [02:12<00:00, 240.72it/s]
该数据集有0/31925张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_288_75/noann_result中以使进行查看
该数据集有0/31925张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_288_75/notrgb_result中以使进行查看
数据集处理完成用时2.22分
loading annotations into memory...
Done (t=0.67s)
loading annotations into memory...
Done (t=0.14s)
creating index...
index created!
100%|██████████| 28577/28577 [02:28<00:00, 192.43it/s]
该数据集有0/28577张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_320_75/noann_result中以使进行查看
该数据集有0/28577张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_320_75/notrgb_result中以使进行查看
数据集处理完成用时2.49分
loading annotations into memory...
Done (t=0.71s)
loading annotations into memory...
Done (t=0.15s)
creating index...
index created!
100%|██████████| 29367/29367 [02:29<00:00, 196.12it/s]
该数据集有0/29367张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_352_75/noann_result中以使进行查看
该数据集有0/29367张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_352_75/notrgb_result中以使进行查看
数据集处理完成用时2.51分
loading annotations into memory...
Done (t=0.67s)
loading annotations into memory...
Done (t=0.13s)
creating index...
index created!
100%|██████████| 24956/24956 [02:29<00:00, 167.27it/s]
该数据集有0/24956张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_384_75/noann_result中以使进行查看
该数据集有0/24956张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_384_75/notrgb_result中以使进行查看
数据集处理完成用时2.5分
loading annotations into memory...
Done (t=0.10s)
loading annotations into memory...
Done (t=0.67s)
creating index...
index created!
100%|██████████| 21233/21233 [02:05<00:00, 168.79it/s]
该数据集有0/21233张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_416_75/noann_result中以使进行查看
该数据集有0/21233张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_416_75/notrgb_result中以使进行查看
数据集处理完成用时2.11分
loading annotations into memory...
Done (t=0.11s)
loading annotations into memory...
Done (t=0.12s)
creating index...
index created!
100%|██████████| 21508/21508 [02:34<00:00, 139.29it/s]
该数据集有0/21508张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_448_75/noann_result中以使进行查看
该数据集有0/21508张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_448_75/notrgb_result中以使进行查看
数据集处理完成用时2.59分
loading annotations into memory...
Done (t=0.08s)
loading annotations into memory...
Done (t=0.09s)
creating index...
index created!
100%|██████████| 16187/16187 [02:15<00:00, 119.87it/s]
该数据集有0/16187张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_480_75/noann_result中以使进行查看
该数据集有0/16187张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_480_75/notrgb_result中以使进行查看
数据集处理完成用时2.25分
loading annotations into memory...
Done (t=0.08s)
loading annotations into memory...
Done (t=0.09s)
creating index...
index created!
100%|██████████| 15916/15916 [02:17<00:00, 115.38it/s]
该数据集有0/15916张图片没有instance标注信息，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_512_75/noann_result中以使进行查看
该数据集有0/15916张图片是非RGB图像，已经这些图片复制到dataset/pqdetection_sliced2voc/pq_512_75/notrgb_result中以使进行查看
数据集处理完成用时2.31分
! tree -L 2 dataset/pqdetection_sliced2voc
[01;34mdataset/pqdetection_sliced2voc[0m
├── [01;34mpq_160_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_192_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_224_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_256_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_288_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_320_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_352_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_384_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_416_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_448_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_480_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
└── [01;34mpq_512_75[0m
├── [01;34mimages[0m
└── [01;34mlabels[0m
37 directories, 0 files
对数据进行可视化，最方便的就是把数据用labelimg查看，如果是在远程服务器上，可以把数据下载下来或vnc远程桌而查看;也可以参考https://blog.csdn.net/u011119817/article/details/119004535这个博客里
在jupyter中显示图片，我们用这种方法
import os
import numpy as np
import matplotlib.pyplot as plt
import xml.etree.ElementTree as ET
def draw_single_image(imgpath,annpath,savepath):
"""_summary_
Args:
imgpath (_type_): 图片路径
annpath (_type_): 标签文件
savepath (_type_): 保存位置
"""
img = cv2.imdecode(np.fromfile(imgpath,dtype=np.uint8),1) #bgr
if img is None or not img.any():
raise Exceeption("read img failded!")
tree = ET.parse(annpath)
root = tree.getroot()
result = root.findall("object")
for obj in result:
name = obj.find("name").text
x1 = int(obj.find("bndbox").find("xmin").text)
y1 = int(obj.find("bndbox").find("ymin").text)
x2 = int(obj.find("bndbox").find("xmax").text)
y2 = int(obj.find("bndbox").find("ymax").text)
cv2.rectangle(img,(x1,y1),(x2,y2),(0,0,255),2)
if savepath is None:
imgrgb = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
plt.figure(figsize=(20,10))
plt.imshow(imgrgb)
else:
cv2.imencode('.jpg',img)[1].tofile(save_path)
draw_single_image(imgpath="dataset/pqdetection_sliced2voc/pq_160_75/images/1_280_120_440_280.jpg", annpath="dataset/pqdetection_sliced2voc/pq_160_75/labels/1_280_120_440_280.xml", savepath=None)
draw_single_image(imgpath="dataset/pqdetection_sliced2voc/pq_224_75/images/2023_0302_184834_009_3270_392_392_616_616.jpg", annpath="dataset/pqdetection_sliced2voc/pq_224_75/labels/2023_0302_184834_009_3270_392_392_616_616.xml", savepath=None)
看到一张没有问题，大概率我们的程序处理流程就没有问题，可以放心往下走
4、生成训练数据
paddledetection训练数据主要包括三个文件train.txt,val.txt,label_list.txt
而我们有两个尺度，所以会有两份的train.txt和val.txt，因为是同一个标签，所以label_list.txt就是ball
with open("dataset/pqdetection_sliced2voc/label_list.txt","w") as f:
f.writelines(["ball"])
# 416的尺度
size_416=[320,352,384,416,448,480,512]
# 256的尺度
size_256=[160,192,224,256,288,320,352]
import os
import random
img_xmls=[]
for i in size_416:
imgpath = f"dataset/pqdetection_sliced2voc/pq_{i}_75/images"
labpath = f"dataset/pqdetection_sliced2voc/pq_{i}_75/labels"
imgs = os.listdir(imgpath)
for img in imgs:
imgbasename = os.path.splitext(img)[0]
imgname = os.path.join(f"pq_{i}_75/images",img)
xmlname = os.path.join(f"pq_{i}_75/labels",imgbasename+'.xml')
img_xml = f"{imgname} {xmlname}\n"
img_xmls.append(img_xml)
for i in range(5):
random.shuffle(img_xmls)
with open("dataset/pqdetection_sliced2voc/train_416.txt","w") as f:
f.writelines(img_xmls)
img_xmls=[]
for i in size_256:
imgpath = f"dataset/pqdetection_sliced2voc/pq_{i}_75/images"
labpath = f"dataset/pqdetection_sliced2voc/pq_{i}_75/labels"
imgs = os.listdir(imgpath)
for img in imgs:
imgbasename = os.path.splitext(img)[0]
imgname = os.path.join(f"pq_{i}_75/images",img)
xmlname = os.path.join(f"pq_{i}_75/labels",imgbasename+'.xml')
img_xml = f"{imgname} {xmlname}\n"
img_xmls.append(img_xml)
for i in range(5):
random.shuffle(img_xmls)
with open("dataset/pqdetection_sliced2voc/train_256.txt","w") as f:
f.writelines(img_xmls)
#看几行训练文件
!head dataset/pqdetection_sliced2voc/train_256.txt
#总的训练样本数
!cat dataset/pqdetection_sliced2voc/train_256.txt | wc -l
pq_192_75/images/2023_0302_175706_008_240_384_432_576_624.jpg pq_192_75/labels/2023_0302_175706_008_240_384_432_576_624.xml
pq_288_75/images/2023_0302_164151_003_780_144_216_432_504.jpg pq_288_75/labels/2023_0302_164151_003_780_144_216_432_504.xml
pq_224_75/images/pingpang_720_00384_560_224_784_448.jpg pq_224_75/labels/pingpang_720_00384_560_224_784_448.xml
pq_160_75/images/2000_0107_022408_002_4830_600_520_760_680.jpg pq_160_75/labels/2000_0107_022408_002_4830_600_520_760_680.xml
pq_256_75/images/2023_0302_164336_005_150_448_464_704_720.jpg pq_256_75/labels/2023_0302_164336_005_150_448_464_704_720.xml
pq_320_75/images/pingpang_720_08579_560_80_880_400.jpg pq_320_75/labels/pingpang_720_08579_560_80_880_400.xml
pq_256_75/images/2023_0302_164336_005_1830_192_256_448_512.jpg pq_256_75/labels/2023_0302_164336_005_1830_192_256_448_512.xml
pq_320_75/images/2023_0302_175706_008_1860_480_160_800_480.jpg pq_320_75/labels/2023_0302_175706_008_1860_480_160_800_480.xml
pq_224_75/images/2023_0302_164336_005_870_392_336_616_560.jpg pq_224_75/labels/2023_0302_164336_005_870_392_336_616_560.xml
pq_224_75/images/10_112_256_336_480.jpg pq_224_75/labels/10_112_256_336_480.xml
255337
因为训练过程中使用在线数据增强方式，所以test就用训练集做，这可能不科学，也就是说test是增强后的train的子集，但至少能说明模型收敛情况。
!cp dataset/pqdetection_sliced2voc/train_256.txt dataset/pqdetection_sliced2voc/test_256.txt
!cp dataset/pqdetection_sliced2voc/train_416.txt dataset/pqdetection_sliced2voc/test_416.txt
! tree -L 2 dataset/pqdetection_sliced2voc
[01;34mdataset/pqdetection_sliced2voc[0m
├── [01;34mpq_160_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_192_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_224_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_256_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_288_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_320_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_352_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_384_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_416_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_448_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_480_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [01;34mpq_512_75[0m
│ ├── [01;34mimages[0m
│ └── [01;34mlabels[0m
├── [00mtest_256.txt[0m
├── [00mtest_416.txt[0m
├── [00mtrain_256.txt[0m
└── [00mtrain_416.txt[0m
37 directories, 4 files
5、模型训练
以256为例，预训练模型使用我们之前的即可,更多配置说明可以参考 https://blog.csdn.net/u011119817/article/details/126665278
配置文件内容大概是：
_BASE_: [
'../datasets/voc_pingpang_20230405.yml',
'../runtime.yml',
'_base_/picodet_v2.yml',
'_base_/optimizer_300e.yml',
]
#pretrain_weights: pretrained/PPLCNet_x0_75_pretrained.pdparams
pretrain_weights: output/picodet_s_416_voc_npu_256_20230404/model_final.pdparams
weights: output/picodet_s_416_voc_npu_256_20230404/model_final.pdparams
find_unused_parameters: True
keep_best_weight: True
use_ema: True
epoch: 80
snapshot_epoch: 5
PicoDet:
backbone: LCNet
neck: CSPPAN
head: PicoHeadV2
LCNet:
scale: 0.75
feature_maps: [3, 4, 5]
act: relu6
CSPPAN:
out_channels: 96
use_depthwise: True
num_csp_blocks: 1
num_features: 4
act: relu6
PicoHeadV2:
conv_feat:
name: PicoFeat
feat_in: 96
feat_out: 96
num_convs: 4
num_fpn_stride: 4
norm_type: bn
share_cls_reg: True
use_se: True
act: relu6
feat_in_chan: 96
act: relu6
LearningRate:
base_lr: 0.0008
schedulers:
- !CosineDecay
max_epochs: 100
min_lr_ratio: 0.001
last_plateau_epochs: 10
- !ExpWarmup
epochs: 5
worker_num: 6
eval_height: &eval_height 256
eval_width: &eval_width 256
eval_size: &eval_size [*eval_height, *eval_width]
TrainReader:
sample_transforms:
- Decode: {}
- Mosaic:
prob: 0.6
input_dim: [416, 416]
degrees: [-10, 10]
scale: [0.1, 2.0]
shear: [-2, 2]
translate: [-0.1, 0.1]
enable_mixup: True
- AugmentHSV: {is_bgr: False, hgain: 5, sgain: 30, vgain: 30}
- RandomFlip: {prob: 0.5}
batch_transforms:
# - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512], random_size: True, random_interp: True, keep_ratio: False}
- BatchRandomResize: {target_size: [160, 192, 224, 256, 288, 320, 352], random_size: True, random_interp: True, keep_ratio: False}
- NormalizeImage: {mean: [0, 0, 0], std: [1, 1, 1], is_scale: True}
- Permute: {}
- PadGT: {}
batch_size: 80
shuffle: true
drop_last: true
mosaic_epoch: 70
EvalReader:
sample_transforms:
- Decode: {}
- Resize: {interp: 2, target_size: *eval_size, keep_ratio: False}
- NormalizeImage: {mean: [0, 0, 0], std: [1, 1, 1], is_scale: True}
- Permute: {}
batch_transforms:
- PadBatch: {pad_to_stride: 32}
batch_size: 16
shuffle: false
TestReader:
inputs_def:
image_shape: [1, 3, *eval_height, *eval_width]
sample_transforms:
- Decode: {}
- Resize: {interp: 2, target_size: *eval_size, keep_ratio: False}
- NormalizeImage: {mean: [0, 0, 0], std: [1, 1, 1], is_scale: True}
- Permute: {}
batch_size: 1
voc_pingpang_20230405.yml内容是：
metric: VOC
map_type: 11point
num_classes: 1
TrainDataset:
!VOCDataSet
dataset_dir: dataset/pqdetection_sliced2voc
anno_path: train_256.txt
label_list: label_list.txt
data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']
EvalDataset:
!VOCDataSet
dataset_dir: dataset/pqdetection_sliced2voc
anno_path: test_256.txt
label_list: label_list.txt
data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']
TestDataset:
!ImageFolder
anno_path: dataset/pqdetection/label_list.txt
命令是：
python -m paddle.distributed.launch --gpus 2,3 tools/train.py -c configs/picodet/picodet_s_416_voc_npu_256_20230405.yml --eval
6、模型推理
对于小目标的推理，同样的是使用滑窗的方式，相关命令查看https://blog.csdn.net/u011119817/article/details/125665278 第8小节内容。
扫一扫
2993
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
