Data parsing on Open Images v4 to speed things "https://towardsdatascience.com/faster-r-cnn-object-detection-implemented-by-keras-for-custom-data-from-googles-open-images-125f62b9141a"

parsing data.py

s = class_descriptions[class_descriptions['class']\
                       .isin(classes)]\
                       .set_index('name').T.to_dict()
train_df = pd.DataFrame(columns=['FileName', 'XMin', 'XMax',
                                 'YMin', 'YMax', 
                                 'ClassName'])

# Find boxes in each image and put them in a dataframe
train_imgs = os.listdir(train_path)
train_imgs = [name[0:16] for name in train_imgs\
              if not name.startswith('.')]

df = annotations_bbox[annotations_bbox.ImageID\
                      .isin(train_imgs)]
df = df[df.LabelName.isin(label_names)]
dd = df.groupby("ImageID")
for ii in train_imgs:
    cc = dd.get_group(ii)
    for index, row in cc.iterrows():
        train_df = train_df.append({'FileName': '{}.jpg'\
                                  .format(ii), 
                                  'XMin': row['XMin'], 
                                  'XMax': row['XMax'], 
                                  'YMin': row['YMin'], 
                                  'YMax': row['YMax'], 
                                  'ClassName': \
                                   s[row['LabelName']]['class']}, 
                                  ignore_index=True)