Skip to content

Failed to red image correct #799

@i-find-flaws-in-AI-programming

Description

Bug Report: Visual Misidentification in DeepSeek Image Analysis

Bug ID: DEEPSEEK-VISION-001
Date: 2025-04-09
Reported by: [User]


Description

When analyzing two full-base screenshots from Clash of Clans (IMG_2089.png and IMG_2090.png), the DeepSeek model incorrectly identified an Archer Tower as a Cannon. This led to an erroneous count of 8 Cannons at Town Hall 11, where the actual maximum is 7 Cannons per game data.


Evidence

Images Provided:

  • IMG_2089.png — top half of base
  • IMG_2090.png — bottom half of base

User-Confirmed Facts:

  • Town Hall level: 11 (visible icon confirms)
  • Cannons: All maxed (level 15)

What the Model Did Wrong

  1. Misidentified Structure:

    • In the bottom center of the base (above the "Attack!" button), an Archer Tower was visually mistaken for a Cannon.
  2. Overcounted:

    • Model counted 3 Cannons in the bottom row:
      • Bottom left ✅ (correct)
      • Bottom center ❌ (incorrect — actually an Archer Tower)
      • Bottom right ✅ (correct)
    • Combined with 5 Cannons from the rest of the base, total became 8.
  3. Failed Cross-Check:

    • Did not compare detected count against known game limit (TH11 max Cannons = 7) to flag the discrepancy.

Root Cause

  • Visual Similarity Confusion: The model relied too heavily on general shape (round base structure behind walls) without distinguishing finer architectural differences between Cannons and Archer Towers.
  • Lack of Game-Knowledge Integration: No validation step was applied to ensure the count did not exceed the maximum possible for the given Town Hall level.
  • Context Ignorance: The model did not verify whether the 8th candidate might already be accounted for as another building type (e.g., an Archer Tower already counted elsewhere).

Correct Answer

  • Total Cannons: 7
  • Total Cannons Needing Upgrade: 0 (all maxed)
  • Correct Count Across Base:
    1. Top left
    2. Top right
    3. Left of Town Hall
    4. Right of Town Hall
    5. Behind Clan Castle
    6. Bottom left corner
    7. Bottom right corner

Suggested Fixes

  1. Improve Visual Differentiation Training:

    • Train the model on finer distinguishing features between visually similar defensive buildings (e.g., Cannon vs. Archer Tower).
  2. Implement Game-Knowledge Validation:

    • Cross-reference detected counts against known game data (e.g., "TH11 max Cannons = 7") and flag any result that exceeds the limit for re-review.
  3. Add Multi-Feature Verification:

    • Require at least two distinguishing attributes (e.g., shape + position relative to walls + comparison to nearby structures) before finalizing identification.

Impact

  • User Trust: Repeated errors eroded confidence in the model's ability to accurately read and analyze images.
  • Accuracy: Minor but persistent misidentifications lead to incorrect conclusions and user frustration.

Attachments

  • IMG_2089.png
  • IMG_2090.png
  • Bug report code (provided separately)

Report submitted by: [I find flaws in AI programming

Image Image Image Image Image

For: DeepSeek Vision Development Team

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions