Confusion in dataset's data_captions.json

[MarketaJu]

I downloaded RSTPReid dataset using provided Google Drive link. However, in my opinion, data_captions.json contains many mistakes.
To specify (since all data are stored as one line, I used img_path for orientation in the data file):

minor issues (can be manually filtered)

• Several Unicode codes for characters used in Chinese, Japanese, and Korean (0001_c5_0025.jpg, 0008_c14_0022.jpg, 0165_c7_0009.jpg, ...)
• Parts of python(?) code (1064_c7_0005.jpg, 2329_c13_0010.jpg, ...)

major issues (can't be manually filtered)

• Some descriptions are identical for multiple images with different id, even though it's clear the text does not match the image. In these cases, the other captions of the images are mostly correct. There is significant amount of these cases.

examples:

• Text description: A man with black hair, wearing glasses, a gray and black shirt, black pants and black shoes, carrying a black backpack, is walking, a hand in the pocket

  • Used for images: 0000_c5_0022.jpg, 0001_c14_0033.jpg, 0001_c1_0014.jpg, 0001_c7_0018.jpg, 0002_c14_0024.jpg, 0002_c1_0003.jpg, 0003_c14_0027.jpg, 0003_c1_0000.jpg, 0005_c14_0030.jpg, i.e. 5 different persons
  • On the images 0002*.jpg, there is clearly long-haired woman in blue jeans with a crossbody bag

• Text description: A woman in a white blouse, black pants and white shoes, hands in pockets, was walking, leading the bag quicklyt

  • Used for images: 0031_c14_0014.jpg, 0031_c1_0003.jpg, 0031_c5_0009.jpg, 0036_c14_0018.jpg, 0036_c1_0002.jpg, 0036_c5_0016.jpg, 0036_c7_0011.jpg
  • In images 0031*.jpg, there is lady fully dressed in black witch colored scarf (for example the second caption: A woman with short hair was wearing sunglasses and a colored scarf, a black coat, black trousers and black flat shoes. She walked down the street with her bag in her right hand. )
  • In images 0036*.jpg, there is lady fully dressed in black with blue backpack and red bag (for example the second caption: A woman with curly hair is wearing a black coat, black trousers and black shoes. She is walking in the street with a blue backpack on her back and a red bag in her left hand.)

Has anyone else encountered similar issues with the dataset, particularly with mismatched captions for different id? If so, it would be great to know how you handled these inconsistencies. Could there be a mistake in how the data was annotated, or is this a known issue?

[792864625]

I also noticed this issue. I filtered that with CLIP, and it looks work.