Update README.md
Browse files
README.md
CHANGED
@@ -1,6 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
π 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 |
π’ Categories of Digit Anomalies
|
5 |
πΈ Digit shape confusion (similar glyphs) β 11 β 17, 21 β 27, 71 β 77
|
6 |
π Mirror / rotation confusion β 69 β 96, 68 β 86, 89β98, 26 β 62
|
@@ -8,18 +19,26 @@
|
|
8 |
π Closed vs. open loop confusion β 38 β 88, 98 β 99, 18 β 19, 56β58, 28β88
|
9 |
βΏ Nearly identical when repeated β 88 β 89, 11 β 12, 55 β 56
|
10 |
π Human OCR-like errors (CAPTCHA/OCR cases) β 47 β 17, 57 β 37, 12 β 72, 14 β 74
|
|
|
|
|
|
|
11 |
π― Applications
|
12 |
π§ͺ Benchmarking OCR systems
|
13 |
π‘ Studying digit recognition robustness
|
14 |
π Training models for noisy / CAPTCHA-like digits
|
15 |
π¨ Anomaly detection in digit datasets
|
|
|
|
|
16 |
βοΈ Technical Details
|
17 |
π Total images: 23,000
|
18 |
π Categories: 23 confusion pairs
|
19 |
βοΈ Font: Humor-Sans.ttf
|
20 |
π Font size: 32
|
21 |
π Image cell size: 60 Γ 40 pixels, 2400x1000 image resolution
|
|
|
22 |
π This dataset provides a controlled testbed for studying digit misclassification under visually ambiguous conditions.
|
|
|
|
|
23 |
π¦ How to Use
|
24 |
1οΈβ£ JSONL format (VQA-style for VLM testing) (merged_puzzles.jsonl)
|
25 |
Each entry includes:
|
@@ -31,11 +50,15 @@
|
|
31 |
πΌ image β file path
|
32 |
π location β anomaly position (row, col)
|
33 |
merged_puzzles.zip file contains all the images.
|
|
|
|
|
34 |
π Suggested Use Cases
|
35 |
π€ VLM evaluation β Test Qwen-VL, InternVL, LLaVA on fine-grained OCR tasks
|
36 |
π OCR benchmarking β Compare CNN-based OCR vs. multimodal LLMs
|
37 |
π Data augmentation research β Train models to handle ambiguity
|
38 |
π΅οΈ Anomaly detection β Use confusion pairs as βhard negativesβ for OCR
|
|
|
|
|
39 |
π§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 |
πΌ 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 |
π 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 |
π’ Categories of Digit Anomalies
|
16 |
πΈ Digit shape confusion (similar glyphs) β 11 β 17, 21 β 27, 71 β 77
|
17 |
π Mirror / rotation confusion β 69 β 96, 68 β 86, 89β98, 26 β 62
|
|
|
19 |
π Closed vs. open loop confusion β 38 β 88, 98 β 99, 18 β 19, 56β58, 28β88
|
20 |
βΏ Nearly identical when repeated β 88 β 89, 11 β 12, 55 β 56
|
21 |
π Human OCR-like errors (CAPTCHA/OCR cases) β 47 β 17, 57 β 37, 12 β 72, 14 β 74
|
22 |
+
|
23 |
+
|
24 |
+
|
25 |
π― Applications
|
26 |
π§ͺ Benchmarking OCR systems
|
27 |
π‘ Studying digit recognition robustness
|
28 |
π Training models for noisy / CAPTCHA-like digits
|
29 |
π¨ Anomaly detection in digit datasets
|
30 |
+
|
31 |
+
|
32 |
βοΈ Technical Details
|
33 |
π Total images: 23,000
|
34 |
π Categories: 23 confusion pairs
|
35 |
βοΈ Font: Humor-Sans.ttf
|
36 |
π Font size: 32
|
37 |
π Image cell size: 60 Γ 40 pixels, 2400x1000 image resolution
|
38 |
+
|
39 |
π This dataset provides a controlled testbed for studying digit misclassification under visually ambiguous conditions.
|
40 |
+
|
41 |
+
|
42 |
π¦ How to Use
|
43 |
1οΈβ£ JSONL format (VQA-style for VLM testing) (merged_puzzles.jsonl)
|
44 |
Each entry includes:
|
|
|
50 |
πΌ image β file path
|
51 |
π location β anomaly position (row, col)
|
52 |
merged_puzzles.zip file contains all the images.
|
53 |
+
|
54 |
+
|
55 |
π Suggested Use Cases
|
56 |
π€ VLM evaluation β Test Qwen-VL, InternVL, LLaVA on fine-grained OCR tasks
|
57 |
π OCR benchmarking β Compare CNN-based OCR vs. multimodal LLMs
|
58 |
π Data augmentation research β Train models to handle ambiguity
|
59 |
π΅οΈ Anomaly detection β Use confusion pairs as βhard negativesβ for OCR
|
60 |
+
|
61 |
+
|
62 |
π§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 |
πΌ Native-resolution ViT (NaViT) β preserves fine details for loop/ stroke differences
|