使用 geopandas(或其他工具)从庞大的数据集中查找嵌套框

find nested boxes from huge dataset with geopandas (or other tools)

基本上我 DataFrame 有大量用 xmin ymin xmax ymax 元组定义的框。

    xmin    ymin    xmax    ymax
0   66      88      130     151
1   143     390     236     468
2   77      331     143     423
3   289     112     337     157
4   343     282     405     352
.....

我的任务是删除所有嵌套框。 (即必须删除 within 另一个框的任何框)

我目前的方法:

沙盒:https://www.kaggle.com/code/easzil/remove-nested-bbox/

def remove_nested_bbox(df):
  # make an extra unique 'id'
  df['__id'] = range(0, len(df))
  # create geometry
  df['__geometry'] = df.apply(lambda x: shapely.geometry.box(x.xmin, x.ymin, x.xmax, x.ymax), axis=1)
  gdf = gpd.GeoDataFrame(df, geometry='__geometry')
  # sort by area
  gdf['__area'] = gdf.__geometry.area
  gdf.sort_values('__area', ascending=False, inplace=True)

  nested_id = set()
  for iloc in range(len(gdf)):
    # skip aready identifed one
    if gdf.iloc[iloc]['__id'] in nested_id:
      continue
    bbox  = gdf.iloc[iloc]    # current target larger bbox
    tests = gdf.iloc[iloc+1:] # all bboxes smaller than the urrent target 
    tests = tests[~tests['__id'].isin(nested_id)] # skip aready identifed one
    nested = tests[tests['__geometry'].within(bbox['__geometry'])]
    nested_id.update(list(nested['__id']))
    
  df = df[~df['__id'].isin(nested_id)]
  
  del df['__id']
  del df['__geometry']
  del df['__area']
    
  return df

有没有更好的方法来优化任务以使其更快? 目前的方法处理大型数据集的速度很慢。

我也会考虑其他方法,例如用 C 或 CUDA 实现。

  • 您的示例数据不大,并且没有框内框的实例。随机生成了一些
  • 已使用 loc 检查维度更大的方法
  • 不确定这是否比您的方法更快,时间细节
%timeit gdf["within"] = gdf.apply(within, args=(gdf,), axis=1)
print(f"""number of polygons: {len(gdf)}
number kept: {len(gdf.loc[lambda d: ~d["within"]])}
""")
2.37 s ± 118 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
number of polygons: 2503
number kept: 241

视觉效果

完整代码

import pandas as pd
import numpy as np
import geopandas as gpd
import io
import shapely

df = pd.read_csv(
    io.StringIO(
        """    xmin    ymin    xmax    ymax
0   66      88      130     151
1   143     390     236     468
2   77      331     143     423
3   289     112     337     157
4   343     282     405     352"""
    ),
    sep="\s+",
)

# randomly generate some boxes, check they are valid
df = pd.DataFrame(
    np.random.randint(1, 200, [10000, 4]), columns=["xmin", "ymin", "xmax", "ymax"]
).loc[lambda d: (d["xmax"] > d["xmin"]) & (d["ymax"] > d["ymin"])]

gdf = gpd.GeoDataFrame(
    df, geometry=df.apply(lambda r: shapely.geometry.box(*r), axis=1)
)

gdf.plot(edgecolor="black", alpha=0.6)

# somewhat optimised by limiting polygons that are considered by looking at dimensions
def within(r, gdf):
    for g in gdf.loc[
        ~(gdf.index == r.name)
        & gdf["xmin"].lt(r["xmin"])
        & gdf["ymin"].lt(r["ymin"])
        & gdf["xmax"].gt(r["xmax"])
        & gdf["ymax"].gt(r["ymax"]),
        "geometry",
    ]:
        if r["geometry"].within(g):
            return True
    return False


gdf["within"] = gdf.apply(within, args=(gdf, ), axis=1)

gdf.loc[lambda d: ~d["within"]].plot(edgecolor="black", alpha=0.6)

方法 2

  • 使用您在 kaggle 上提供的样本数据
  • 这个 returns 与以前的版本相比,用了大约一半的时间 (5s)
  • 概念类似,如果 xmin 和 ymin 大于另一个框且 max 和 ymax 小于另一个框,则框在另一个框内
import functools

df = pd.read_csv("https://storage.googleapis.com/kagglesdsdata/datasets/2015126/3336994/sample.csv?X-Goog-Algorithm=GOOG4-RSA-SHA256&X-Goog-Credential=gcp-kaggle-com%40kaggle-161607.iam.gserviceaccount.com%2F20220322%2Fauto%2Fstorage%2Fgoog4_request&X-Goog-Date=20220322T093633Z&X-Goog-Expires=259199&X-Goog-SignedHeaders=host&X-Goog-Signature=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")

def within_np(df):
    d = {}
    for c in df.columns[0:4]:
        a = np.tile(df[c].values.T ,(len(df),1))
        d[c] = a.T > a if c[1:] == "min" else a.T < a

    aa = functools.reduce(np.logical_and, (aa for aa in d.values()))
    return aa.sum(axis=1)>0

df.loc[~within_np(df)]