ImagenetV2 has a lot of incorrect Bbox annotation information?

[fengxun2017]

I used the bBox annotation information provided by you for testing, and found many wrong Bboxes. for example: 100/5.jpeg,73,214,148,299 100/5.jpeg,180,85,200,111 100/5.jpeg,169,137,231,195 100/5.jpeg,394,163,462,207 100/5.jpeg,316,134,374,160 100/5.jpeg,136,206,171,235 100/5.jpeg,242,146,267,164

[junsukchoe]

Hello,

Thanks for your interest in our work! I just checked 100/5.jpeg, but I cannot find any wrong box:

Please check again and let me know if you find anything wrong from the box annotations.

[huang422]

I cropped all the bbox and also find a lot of incorrect Bbox annotation? val2/2/8.jpeg,173,109,462,224 val2/58/5.jpeg,112,220,349,310 val2/35/6.jpeg,36,38,331,232 val2/36/1.jpeg,72,73,451,277 val2/48/9.jpeg,0,28,420,257 ... ... Thanks

[ShaileshSridhar2403]

Hi @junsukchoe , I also encounter the same problem, here are the bounding boxes on 100/5.jpeg visualized

Bounding boxes as provided here are:

38,41,66,68 73,214,148,299 180,85,200,111 169,137,231,195 394,163,462,207 316,134,374,160 136,206,171,235 242,146,267,164

How did you visualize the bboxes in your comment above? Is there a step we are missing?

Here is another example : 101/8.jpeg

[junsukchoe]

Hi @ShaileshSridhar2403, @fengxun2017!

I just reproduced the above problems in the ImageNetV2 dataset. The cause of the problem is that the actual image size is different from the size mentioned in our 'image_sizes.txt'. For example, the size of 100/5.jpeg (actually it is now 100/1f3075074e2c005496d52faa07089f3cea130dee.jpeg) is (384, 256), while the size listed in images_sizes.txt is (500, 333). Hence, you have to first resize the image to (500, 333) and then draw the bboxes for the correct qualitative samples.

Here is an example code snippet:

import cv2
import matplotlib.pyplot as plt

img = cv2.cvtColor(cv2.imread('100/5.jpeg'), cv2.COLOR_BGR2RGB)
img = cv2.resize(img, dsize=(500,333))
img = cv2.rectangle(img, (38, 41), (66, 68), (255,0,0), 3)
img = cv2.rectangle(img, (73, 214), (148, 299), (255,0,0), 3)
img = cv2.rectangle(img, (180, 85), (200, 111), (255,0,0), 3)
img = cv2.rectangle(img, (169, 137), (231, 195), (255,0,0), 3)
img = cv2.rectangle(img, (394, 163), (462, 207), (255,0,0), 3)
img = cv2.rectangle(img, (316, 134), (374, 160), (255,0,0), 3)
img = cv2.rectangle(img, (136, 206), (171, 235), (255,0,0), 3)
img = cv2.rectangle(img, (242, 146), (267, 164), (255,0,0), 3)

plt.imshow(img)

Then, you will get the following image:

Unfortunately, it is currently unknown why this discripancy happened. But I would like to note that the evaluation results from our code are still correct since we convert the bboxes to relative coordinates for evaluation using the following function: https://github.com/clovaai/wsolevaluation/blob/e00842f8e9d86588d45f8e3e30c237abb364bba4/evaluation.py#L95

I am sincerly sorry for the late reply. Please let me know if you have any further questions.