what labeling format has negative Bbox values in labels?

Question

I have a labeled dataset for object detection few thousands of images with annotation on csv file

the csv contains these columns image_path, class, xmax, xmin, ymax, ymin

looks like Pascal voc format labeles but when I plot it, it does't work and all labels look wrong

I noticed that there are some values in negtive in xmin, ymin and that's weird

so I'm asking which kind of labeling formats that has the possibility to have these kinds of values because it's ?

as example a bbox has these

xmax: 429 xmin: -68 ymax = 390 ymin = 130

the code I used is:

import cv2
import pandas as pd 
df = pd.read_csv('train.csv')

directory = 'dataset/images'

color = (255, 0, 0)
thickness = 2
org = (50, 50)
fontScale = 1
font = cv2.FONT_HERSHEY_SIMPLEX

for i in df.index:
  image_path = df.iloc[i]['image_path']
  xmax = int(df.iloc[i]['xmax'])
  xmin = int(df.iloc[i]['xmin'])
  ymax = int(df.iloc[i]['ymax'])
  ymin = int(df.iloc[i]['ymin'])
  txt = df.iloc[i]['name']

  x= xmax - xmin
  y = ymax - ymin 
  start_point = (xmin,ymin)
  end_point = (xmax,ymax)

  img = cv2.imread(directory+ '/'+image_path)


  image = cv2.putText(img, txt, org, font, 
               fontScale, color, thickness, cv2.LINE_AA)

  bbox_img = cv2.rectangle(image, start_point, end_point, color, thickness)

  cv2.imshow('image window', bbox_img)
  # add wait key. window waits until user presses a key
  cv2.waitKey(0)
  # and finally destroy/close all open windows
  cv2.destroyAllWindows()

Answer 1

Based on your answer, it looks like some of your bounding boxes use a different coordinate system than what is normally used for plotting in OpenCV. OpenCV treats the coordinates in the image as strictly positive (with $(0,0)$ being the top-left corner on the image when plotted) source.The fact that you have negative coordinates indicates that those coordinates use a different paradigm, e.g. that $(0,0)$ is in the center of the image.

You could look up what the formatting is for your dataset. It also might be simpler to re-normalize the image coordinates so that the lowest $x$ and $y$ values are $(0,0)$