Github仓库：gy-7/coco_EDA (github.com)
对coco数据集的分析，近期忙着写论文，空余时间很少能写博文了。
EDA的代码放在结尾了，Github仓库里也有。仓库里还有其他的一些EDA分析，不定时更新。
训练集所有类别的数量分布情况：
训练集所有类别的尺寸分布情况：
验证集所有类别的数量分布情况：
验证集所有类别的尺寸分布情况：
EDA代码：
import os
import seaborn as sns
import pycocotools.coco
import matplotlib.pyplot as plt
root_dir = os.getcwd()
train_ann_fp = os.path.join(root_dir, 'annotations', 'instances_train2017.json')
val_ann_fp = os.path.join(root_dir, 'annotations', 'instances_val2017.json')
class COCO_EDA:
def __init__(self, json_file, type='train'):
self.COCO_SMALL_SCALE = 32
self.COCO_MEDIUM_SCALE = 96
self.json_file = json_file
coco = pycocotools.coco.COCO(json_file)
self.type = type
self.imgs = coco.dataset['images']
self.anns = coco.dataset['annotations']
self.cats = coco.dataset['categories']
self.img_ids = coco.getImgIds()
self.ann_ids = coco.getAnnIds()
self.cat_ids = coco.getCatIds()
self.cat2imgs = coco.catToImgs
self.img2anns = coco.imgToAnns
self.imgs_num = len(self.imgs)
self.objs_num = len(self.anns)
# data to be collected
self.small_objs_num = 0
self.medium_objss_num = 0
self.large_objss_num = 0
self.small_objs = []
self.medium_objs = []
self.large_objs = []
self.cat2objs = {}
self.small_cat2objs = {} # small objects classes distribution
self.medium_cat2objs = {} # medium objects classes distribution
self.large_cat2objs = {} # large objects classes distribution
self.cat2objs_num = {} # objects classes distribution
self.small_cat2objs_num = {} # small objects classes distribution
self.medium_cat2objs_num = {} # medium objects classes distribution
self.large_cat2objs_num = {} # large objects classes distribution
# plot use data
self.catid2name = {} # 用于绘图中显示类别名字
self.cats_plot = [] # coco 所有尺寸目标的类别分布
self.small_cats_plot = [] # 小目标中每个类的分布情况
self.medium_cats_plot = [] # 中目标中每个类的分布情况
self.large_cats_plot = [] # 大目标中每个类的分布情况
# 每个类的小，中，大目标的数量
self.size_distribution = {}
def collect_data(coco):
# collect small, medium, large objects
for ann in coco.anns:
if ann['area'] < coco.COCO_SMALL_SCALE ** 2:
coco.small_objs_num += 1
coco.small_objs.append(ann)
elif ann['area'] < coco.COCO_MEDIUM_SCALE ** 2:
coco.medium_objs.append(ann)
coco.medium_objss_num += 1
else:
coco.large_objs.append(ann)
coco.large_objss_num += 1
for i in coco.cat_ids:
coco.cat2objs[i] = []
coco.small_cat2objs[i] = []
coco.medium_cat2objs[i] = []
coco.large_cat2objs[i] = []
coco.cat2objs_num[i] = 0
coco.small_cat2objs_num[i] = 0
coco.medium_cat2objs_num[i] = 0
coco.large_cat2objs_num[i] = 0
coco.size_distribution[i] = []
for i in coco.cats:
coco.catid2name[i['id']] = i['name']
# collect small, medium, large class distribution
for i in coco.anns:
coco.cat2objs[i['category_id']].append(i)
coco.cat2objs_num[i['category_id']] += 1
coco.cats_plot.append(coco.catid2name[i['category_id']])
if i['area'] < coco.COCO_SMALL_SCALE ** 2:
coco.small_cat2objs[i['category_id']].append(i)
coco.small_cat2objs_num[i['category_id']] += 1
coco.small_cats_plot.append(coco.catid2name[i['category_id']])
coco.size_distribution[i['category_id']].append('s')
elif i['area'] < coco.COCO_MEDIUM_SCALE ** 2:
coco.medium_cat2objs[i['category_id']].append(i)
coco.medium_cat2objs_num[i['category_id']] += 1
coco.medium_cats_plot.append(coco.catid2name[i['category_id']])
coco.size_distribution[i['category_id']].append('m')
else:
coco.large_cat2objs[i['category_id']].append(i)
coco.large_cat2objs_num[i['category_id']] += 1
coco.large_cats_plot.append(coco.catid2name[i['category_id']])
coco.size_distribution[i['category_id']].append('l')
assert len(coco.small_objs) == coco.small_objs_num == sum(coco.small_cat2objs_num.values())
assert len(coco.medium_objs) == coco.medium_objss_num == sum(coco.medium_cat2objs_num.values())
assert len(coco.large_objs) == coco.large_objss_num == sum(coco.large_cat2objs_num.values())
assert len(coco.anns) == coco.objs_num == sum(coco.cat2objs_num.values())
def plot_coco_class_distribution(plot_data, plot_order, save_fp, plot_title, plot_y_heigh,
plot_y_heigh_residual=[1800, 100]):
# 绘制coco数据集的类别分布
sns.set_style("whitegrid")
plt.figure(figsize=(15, 8)) # 图片的宽和高，单位为inch
plt.title(plot_title, fontsize=9) # 标题
plt.xlabel('class', fontsize=8) # x轴名称
plt.ylabel('counts', fontsize=8) # y轴名称
plt.xticks(rotation=90, fontsize=8) # x轴标签竖着显示
plt.yticks(fontsize=8)
for x, y in enumerate(plot_y_heigh):
if 'train' in save_fp:
plt.text(x, y + plot_y_heigh_residual[0], '%s' % y, ha='center', fontsize=7, rotation=90)
else:
plt.text(x, y + plot_y_heigh_residual[1], '%s' % y, ha='center', fontsize=7, rotation=90)
ax = sns.countplot(x=plot_data, palette="PuBu_r", order=plot_order) # 绘制直方图，palette调色板，蓝色由浅到深渐变。
# palette样式：https://blog.csdn.net/panlb1990/article/details/103851983
plt.savefig(os.path.join(save_fp), dpi=500)
plt.show()
def plot_size_distribution(plot_data, save_fp, plot_title, plot_order=['s', 'm', 'l']):
sns.set_style("whitegrid")
plt.figure(figsize=(21, 35)) # 图片的宽和高，单位为inch
plt.subplots_adjust(left=0.1, bottom=0.1, right=0.9, top=0.9, wspace=1, hspace=1.5) # 调整子图间距
for idx, size_data in enumerate(plot_data.values()):
plt.subplot(10, 8, idx + 1)
plt.xticks(rotation=0, fontsize=18) # x轴标签竖着显示
plt.yticks(fontsize=18)
plt.xlabel('size', fontsize=20) # x轴名称
plt.ylabel('count', fontsize=20) # y轴名称
plt.title(plot_title[idx], fontsize=24) # 标题
sns.countplot(x=size_data, palette="PuBu_r", order=plot_order) # 绘制直方图，palette调色板，蓝色由浅到深渐变。
plt.savefig(save_fp, dpi=500, pad_inches=0)
plt.show()
def run_plot_coco_class_distribution(coco, save_dir):
# # 绘制coco数据集的类别分布
plot_order = [i for i in coco.catid2name.values()]
plot_heigh = [i for i in coco.cat2objs_num.values()]
save_fp = os.path.join(save_dir, f'coco_{coco.type}_class_distribution.png')
plot_coco_class_distribution(coco.cats_plot, plot_order, save_fp, 'COCO train2017 class distribution', plot_heigh,
plot_y_heigh_residual=[1800, 100])
plot_heigh = [i for i in coco.small_cat2objs_num.values()]
save_fp = os.path.join(save_dir, f'coco_{coco.type}_small_class_distribution.png')
plot_coco_class_distribution(coco.small_cats_plot, plot_order, save_fp, 'COCO train2017 small class distribution',
plot_heigh,
plot_y_heigh_residual=[900, 50])
plot_heigh = [i for i in coco.medium_cat2objs_num.values()]
save_fp = os.path.join(save_dir, f'coco_{coco.type}_medium_class_distribution.png')
plot_coco_class_distribution(coco.medium_cats_plot, plot_order, save_fp, 'COCO train2017 medium class distribution',
plot_heigh, plot_y_heigh_residual=[900, 50])
plot_heigh = [i for i in coco.large_cat2objs_num.values()]
save_fp = os.path.join(save_dir, f'coco_{coco.type}_large_class_distribution.png')
plot_coco_class_distribution(coco.large_cats_plot, plot_order, save_fp, 'COCO train2017 large class distribution',
plot_heigh,
plot_y_heigh_residual=[900, 50])
def run_plot_coco_size_distribution(coco, save_dir):
# 绘制coco数据集各类别的尺寸分布
plot_order = [i for i in coco.catid2name.values()]
save_fp = os.path.join(save_dir, f'coco_{coco.type}_size_distribution.png')
plot_size_distribution(coco.size_distribution, save_fp, plot_order)
if __name__ == '__main__':
print("analyze coco train dataset...")
print("-" * 50)
coco_train = COCO_EDA(train_ann_fp, type='train')
collect_data(coco_train)
print("coco train images num:", coco_train.imgs_num)
print("coco train objects num:", coco_train.objs_num)
print("coco small objects num:", coco_train.small_objs_num)
print("coco medium objects num:", coco_train.medium_objss_num)
print("coco large objects num:", coco_train.large_objss_num)
print("coco small objects percent:", coco_train.small_objs_num / coco_train.objs_num)
print("coco medium objects percent:", coco_train.medium_objss_num / coco_train.objs_num)
print("coco large objects percent:", coco_train.large_objss_num / coco_train.objs_num)
run_plot_coco_class_distribution(coco_train, ".\\EDA")
run_plot_coco_size_distribution(coco_train, ".\\EDA")
print("-" * 50)
print()
print("analyze coco val dataset...")
print("-" * 50)
coco_val = COCO_EDA(val_ann_fp, type='val')
collect_data(coco_val)
print("coco val images num:", coco_val.imgs_num)
print("coco val objects num:", coco_val.objs_num)
print("coco small objects num:", coco_val.small_objs_num)
print("coco medium objects num:", coco_val.medium_objss_num)
print("coco large objects num:", coco_val.large_objss_num)
print("coco small objects percent:", coco_val.small_objs_num / coco_val.objs_num)
print("coco medium objects percent:", coco_val.medium_objss_num / coco_val.objs_num)
print("coco large objects percent:", coco_val.large_objss_num / coco_val.objs_num)
run_plot_coco_class_distribution(coco_val, ".\\EDA")
run_plot_coco_size_distribution(coco_val, ".\\EDA")
print("-" * 50)