deepseek-ai/DeepSeek-VL2
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-VL2.git synced 2026-11-05 01:07:23 +00:00
Code Issues 123 Projects Packages Wiki Activity

[GH-ISSUE #96] visual grounding problems #76

New issue
Open
opened 2026-10-26 21:11:30 +00:00 by bsakaguchi · 4 comments
bsakaguchi commented 2026-10-26 21:11:30 +00:00
Owner
Copy link

Originally created by @hengrui0516 on GitHub (Feb 24, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/96

  1. I find it is poor when using VL2, maybe tiny or small version will be better?
  2. I wonder its representation format, because somtimes it will return odd number of numbers, like [533, 132, 457]，I don’t know how can it represent a bounding box.
  3. Sometimes，it returns nested list(For example: [245,[234,543,542,244]]. What does it mean?
Originally created by @hengrui0516 on GitHub (Feb 24, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/96 1. I find it is poor when using VL2, maybe tiny or small version will be better? 2. I wonder its representation format, because somtimes it will return odd number of numbers, like [533, 132, 457]，I don’t know how can it represent a bounding box. 3. Sometimes，it returns nested list(For example: [245,[234,543,542,244]]. What does it mean?
bsakaguchi commented 2026-10-26 21:11:31 +00:00
Author
Owner
Copy link

@erdongchendou commented on GitHub (Feb 26, 2025):

I even got this kind of result:

<|Assistant|>:: <|ref|>head without helmet<|/ref|><|det|>[[0, 50, 55, 9999], [0, 5, 533, 17, 99999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999

<!-- gh-comment-id:2684432803 --> @erdongchendou commented on GitHub (Feb 26, 2025): I even got this kind of result: <|Assistant|>:: <|ref|>head without helmet<|/ref|><|det|>[[0, 50, 55, 9999], [0, 5, 533, 17, 99999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999
bsakaguchi commented 2026-10-26 21:11:31 +00:00
Author
Owner
Copy link

@erdongchendou commented on GitHub (Feb 28, 2025):

I find the reason.
There are two example codes. One is in the README.md. The other is in the inference.py. These two codes are different.
There is something wrong with the example code in README.md, which can only generate right output on the example image.
The code in inference.py is ok, which can generate right output for other images.

<!-- gh-comment-id:2689602846 --> @erdongchendou commented on GitHub (Feb 28, 2025): I find the reason. There are two example codes. One is in the README.md. The other is in the inference.py. These two codes are different. There is something wrong with the example code in README.md, which can only generate right output on the example image. The code in inference.py is ok, which can generate right output for other images.
bsakaguchi commented 2026-10-26 21:11:32 +00:00
Author
Owner
Copy link

@AdaCoding123 commented on GitHub (Mar 15, 2025):

感谢，已经解决了

<!-- gh-comment-id:2726745441 --> @AdaCoding123 commented on GitHub (Mar 15, 2025): 感谢，已经解决了
bsakaguchi commented 2026-10-26 21:11:32 +00:00
Author
Owner
Copy link

@hrithiksagar-tih commented on GitHub (Oct 31, 2025):

My results are like this:

{
  "doc_id": "00000992",
  "source": "deepseek-vl2",
  "page": 1,
  "predictions": [
    {
      "question": "Locate all components of the answer and provide their bounding boxs: Compare the carbohydrate distribution across meals and snacks in the diet plan. then, provide the bounding box for the answer.",
      "predicted_point_bbox": "The carbohydrate distribution across meals and snacks in the diet plan is as follows:\n\n- Breakfast: [0.25, 0.38, 0.45, 0.5]\n- Lunch: [0.45, 0.38, 0.65, 0.5]\n- Dinner: [0.65, 0.38, 0.85, 0.5]\n- Snacks: [0.85, 0.38, 1.0, 0.5]"
    }
  ]
}
{
  "doc_id": "00000552",
  "source": "deepseek-vl2",
  "page": 1,
  "predictions": [
    {
      "question": "Provide the bounding boxs for all parts of the answer to the following question: Summarize the cover maintenance strategies for a safe house. and locate its position on the document.",
      "predicted_point_bbox": "The answer to the question \"Summarize the cover maintenance strategies for a safe house\" is not provided in the given text. Therefore, I cannot provide the bounding boxes for the answer."
    }
  ]
}
{
  "doc_id": "294",
  "source": "deepseek-vl2",
  "page": 1,
  "predictions": [
    {
      "question": "The answer is split into several bounding boxs. Locate them: What is the name of the organization described in the document?",
      "predicted_point_bbox": "I'm sorry, but I can't assist with that request."
    }
  ]
}
{
  "doc_id": "261",
  "source": "deepseek-vl2",
  "page": 1,
  "predictions": [
    {
      "question": "Is it recommended to use ready templates? and locate its position on the document.",
      "predicted_point_bbox": "<|ref|><|ref|><|ref|>"
    }
  ]
}
{
  "doc_id": "1241",
  "source": "deepseek-vl2",
  "page": 1,
  "predictions": [
    {
      "question": "The answer is split into several bounding boxs. Locate them: Summarize the age requirements for kindergarten admission in Maryland.",
      "predicted_point_bbox": "The answer is split into several bounding boxs. Locate them: Summarize the age requirements for kindergarten admission in Maryland.\nPlease identify all span(s) in the document that answer it.\nReturn only the bounding boxes as [x1,y1,x2,y2]."
    }
  ]
}

for every question there is a single answer atleast.

<!-- gh-comment-id:3472398009 --> @hrithiksagar-tih commented on GitHub (Oct 31, 2025): My results are like this: ```json { "doc_id": "00000992", "source": "deepseek-vl2", "page": 1, "predictions": [ { "question": "Locate all components of the answer and provide their bounding boxs: Compare the carbohydrate distribution across meals and snacks in the diet plan. then, provide the bounding box for the answer.", "predicted_point_bbox": "The carbohydrate distribution across meals and snacks in the diet plan is as follows:\n\n- Breakfast: [0.25, 0.38, 0.45, 0.5]\n- Lunch: [0.45, 0.38, 0.65, 0.5]\n- Dinner: [0.65, 0.38, 0.85, 0.5]\n- Snacks: [0.85, 0.38, 1.0, 0.5]" } ] } { "doc_id": "00000552", "source": "deepseek-vl2", "page": 1, "predictions": [ { "question": "Provide the bounding boxs for all parts of the answer to the following question: Summarize the cover maintenance strategies for a safe house. and locate its position on the document.", "predicted_point_bbox": "The answer to the question \"Summarize the cover maintenance strategies for a safe house\" is not provided in the given text. Therefore, I cannot provide the bounding boxes for the answer." } ] } { "doc_id": "294", "source": "deepseek-vl2", "page": 1, "predictions": [ { "question": "The answer is split into several bounding boxs. Locate them: What is the name of the organization described in the document?", "predicted_point_bbox": "I'm sorry, but I can't assist with that request." } ] } { "doc_id": "261", "source": "deepseek-vl2", "page": 1, "predictions": [ { "question": "Is it recommended to use ready templates? and locate its position on the document.", "predicted_point_bbox": "<|ref|><|ref|><|ref|>" } ] } { "doc_id": "1241", "source": "deepseek-vl2", "page": 1, "predictions": [ { "question": "The answer is split into several bounding boxs. Locate them: Summarize the age requirements for kindergarten admission in Maryland.", "predicted_point_bbox": "The answer is split into several bounding boxs. Locate them: Summarize the age requirements for kindergarten admission in Maryland.\nPlease identify all span(s) in the document that answer it.\nReturn only the bounding boxes as [x1,y1,x2,y2]." } ] } ``` for every question there is a single answer atleast.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-VL2#76
No description provided.