Add preprocessing pipeline for NavOCR dataset

[exceldra5]

Goal

NavOCR preprocessing code

• CLIP based exterior filtering
• OCR every text in filtered text
• Match text with levenshtein distance (utilize DeepL API)

Environment

Branch / Commit: preprocessor_v2
Language / Framework: python
Related Branches: crawl_v2, backup-private-ver

Code Update

• backup-private-ver based code update
• use CLIP prompt in backup-private-ver branch
• Write the executable code using the console_scripts approach in the main branch.
• Configure additional dependencies using extras_require in setup.py.

# 1. Install extras for preprocessing and translation
pip install -e ".[preprocess,translate]"

# 2. Install PaddleOCR (manual step — difficult to automate due to CUDA variations)
pip install paddlepaddle paddleocr        # Apple Silicon / CPU
# or
pip install paddlepaddle-gpu paddleocr    # NVIDIA GPU

• env check code

python scripts/check_preprocess_env.py

---

Paper Feedback

No.	Item	Paper §3.1	Code
1	CLIP prompt	Single prompt: "an exterior photo of the place or a photo that includes a signboard"	3-pass cascade using 3 different prompt sets
2	CLIP threshold semantics	Single threshold: 0.7	3-pass × 0.7 (AND semantics)
3	Levenshtein threshold	0.5	0.5
4	DeepL translation fallback	None	Re-matching after Korean↔English translation
5	Number of images/queries	20 images	Using 30 images
6	Label format	("file_name", (x₁,y₁,w₁,h₁), …, (xₙ,yₙ,wₙ,hₙ)) simple tuple	COCO JSON (images, annotations, categories)

• The CLIP prompt was applied in a 3-stage process as described below, using both positive and negative prompts for classification.

PROMPT_SETS: list[tuple[str, str]] = [
	(
		"a photo that are not contain signboard or a photo of a signboard against a plain background",
		"a photo of a store exterior",
	),
	(
		"a photo that are not contain signboard or a photo of a signboard against a plain background",
		"a photo of a store exterior with a signboard and surrounding environment",
	),
	(
		"a online logo of a brand",
		"a photo of a store exterior with a signboard and surrounding environment",
	),
]

• Considering whether DeepL-based translation is necessary.
• In Figure 2, changed the 1st filter from CLIP to OCR, and the 2nd filter to OCR accordingly.
• Changed the label format from a simple tuple (as in the paper) to the provided COCO JSON-based format.

ToDo

• Enhanced code annotations.
• Conducted additional testing.
• Implemented web crawling code.
• Incorporated licensing considerations during crawling.