chandrabhuma commited on
Commit
76caf1a
Β·
verified Β·
1 Parent(s): bc955ee

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +23 -0
README.md CHANGED
@@ -1,6 +1,17 @@
 
 
 
 
 
 
 
 
 
1
  πŸ“Š DigitConfuse-23k: A Synthetic Dataset of Digit Confusion Patterns
2
  ...DigitConfuse-23k is a synthetic dataset containing 23,000 images of digit pairs designed to capture visual anomalies and confusion cases commonly encountered in OCR, CAPTCHA recognition, optical illusions and human digit interpretation tasks.
3
  ...Each image contains two-digit numbers generated using the Humor-Sans font (font_size=32, cell_w=60, cell_h=40). For each confusion category, ~1000 images are included.
 
 
4
  πŸ”’ Categories of Digit Anomalies
5
  πŸ”Έ Digit shape confusion (similar glyphs) β†’ 11 ↔ 17, 21 ↔ 27, 71 ↔ 77
6
  πŸ”„ Mirror / rotation confusion β†’ 69 ↔ 96, 68 ↔ 86, 89↔98, 26 ↔ 62
@@ -8,18 +19,26 @@
8
  πŸŒ€ Closed vs. open loop confusion β†’ 38 ↔ 88, 98 ↔ 99, 18 ↔ 19, 56↔58, 28↔88
9
  ➿ Nearly identical when repeated β†’ 88 ↔ 89, 11 ↔ 12, 55 ↔ 56
10
  πŸ‘€ Human OCR-like errors (CAPTCHA/OCR cases) β†’ 47 ↔ 17, 57 ↔ 37, 12 ↔ 72, 14 ↔ 74
 
 
 
11
  🎯 Applications
12
  πŸ§ͺ Benchmarking OCR systems
13
  πŸ›‘ Studying digit recognition robustness
14
  πŸ”‘ Training models for noisy / CAPTCHA-like digits
15
  🚨 Anomaly detection in digit datasets
 
 
16
  βš™οΈ Technical Details
17
  πŸ“‚ Total images: 23,000
18
  πŸ“‘ Categories: 23 confusion pairs
19
  ✍️ Font: Humor-Sans.ttf
20
  πŸ”  Font size: 32
21
  πŸ“ Image cell size: 60 Γ— 40 pixels, 2400x1000 image resolution
 
22
  πŸ‘‰ This dataset provides a controlled testbed for studying digit misclassification under visually ambiguous conditions.
 
 
23
  πŸ“¦ How to Use
24
  1️⃣ JSONL format (VQA-style for VLM testing) (merged_puzzles.jsonl)
25
  Each entry includes:
@@ -31,11 +50,15 @@
31
  πŸ–Ό image β†’ file path
32
  πŸ“Œ location β†’ anomaly position (row, col)
33
  merged_puzzles.zip file contains all the images.
 
 
34
  πŸš€ Suggested Use Cases
35
  πŸ€– VLM evaluation β†’ Test Qwen-VL, InternVL, LLaVA on fine-grained OCR tasks
36
  πŸ“Š OCR benchmarking β†’ Compare CNN-based OCR vs. multimodal LLMs
37
  πŸ”„ Data augmentation research β†’ Train models to handle ambiguity
38
  πŸ•΅οΈ Anomaly detection β†’ Use confusion pairs as β€œhard negatives” for OCR
 
 
39
  πŸ§ͺ Real-World Testing with Ovis 2.5-9B (Latest Release)
40
  I evaluated a subset of images using Ovis 2.5-9B (released Aug 2025).
41
  πŸ–Ό Native-resolution ViT (NaViT) β†’ preserves fine details for loop/ stroke differences
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - image-text-to-text
5
+ language:
6
+ - en
7
+ size_categories:
8
+ - 10K<n<100K
9
+ ---
10
  πŸ“Š DigitConfuse-23k: A Synthetic Dataset of Digit Confusion Patterns
11
  ...DigitConfuse-23k is a synthetic dataset containing 23,000 images of digit pairs designed to capture visual anomalies and confusion cases commonly encountered in OCR, CAPTCHA recognition, optical illusions and human digit interpretation tasks.
12
  ...Each image contains two-digit numbers generated using the Humor-Sans font (font_size=32, cell_w=60, cell_h=40). For each confusion category, ~1000 images are included.
13
+
14
+
15
  πŸ”’ Categories of Digit Anomalies
16
  πŸ”Έ Digit shape confusion (similar glyphs) β†’ 11 ↔ 17, 21 ↔ 27, 71 ↔ 77
17
  πŸ”„ Mirror / rotation confusion β†’ 69 ↔ 96, 68 ↔ 86, 89↔98, 26 ↔ 62
 
19
  πŸŒ€ Closed vs. open loop confusion β†’ 38 ↔ 88, 98 ↔ 99, 18 ↔ 19, 56↔58, 28↔88
20
  ➿ Nearly identical when repeated β†’ 88 ↔ 89, 11 ↔ 12, 55 ↔ 56
21
  πŸ‘€ Human OCR-like errors (CAPTCHA/OCR cases) β†’ 47 ↔ 17, 57 ↔ 37, 12 ↔ 72, 14 ↔ 74
22
+
23
+
24
+
25
  🎯 Applications
26
  πŸ§ͺ Benchmarking OCR systems
27
  πŸ›‘ Studying digit recognition robustness
28
  πŸ”‘ Training models for noisy / CAPTCHA-like digits
29
  🚨 Anomaly detection in digit datasets
30
+
31
+
32
  βš™οΈ Technical Details
33
  πŸ“‚ Total images: 23,000
34
  πŸ“‘ Categories: 23 confusion pairs
35
  ✍️ Font: Humor-Sans.ttf
36
  πŸ”  Font size: 32
37
  πŸ“ Image cell size: 60 Γ— 40 pixels, 2400x1000 image resolution
38
+
39
  πŸ‘‰ This dataset provides a controlled testbed for studying digit misclassification under visually ambiguous conditions.
40
+
41
+
42
  πŸ“¦ How to Use
43
  1️⃣ JSONL format (VQA-style for VLM testing) (merged_puzzles.jsonl)
44
  Each entry includes:
 
50
  πŸ–Ό image β†’ file path
51
  πŸ“Œ location β†’ anomaly position (row, col)
52
  merged_puzzles.zip file contains all the images.
53
+
54
+
55
  πŸš€ Suggested Use Cases
56
  πŸ€– VLM evaluation β†’ Test Qwen-VL, InternVL, LLaVA on fine-grained OCR tasks
57
  πŸ“Š OCR benchmarking β†’ Compare CNN-based OCR vs. multimodal LLMs
58
  πŸ”„ Data augmentation research β†’ Train models to handle ambiguity
59
  πŸ•΅οΈ Anomaly detection β†’ Use confusion pairs as β€œhard negatives” for OCR
60
+
61
+
62
  πŸ§ͺ Real-World Testing with Ovis 2.5-9B (Latest Release)
63
  I evaluated a subset of images using Ovis 2.5-9B (released Aug 2025).
64
  πŸ–Ό Native-resolution ViT (NaViT) β†’ preserves fine details for loop/ stroke differences