Backend Sign Processing System

This project implements a high-performance computer vision pipeline for detecting signs, stitching panoramic views, and correcting occlusions using multi-view geometry.

Architecture

1. Filtering: Checks input folders (face1...face4).

2. Stitching: If a folder has multiple images, it uses cv2.Stitcher (Scan mode) to create a single high-res panorama.

3. Detection: A placeholder MockSignDetector identifies signs (segmentation masks).

4. Occlusion Recovery:

• If a mask is fragmented (indicating occlusion), the system loads auxiliary images from signface folders.

• It uses ORB Feature Matching and Homography to warp the auxiliary image to the main perspective.

• It checks for overlap and re-runs detection on the clean, warped image to replace the broken mask.

5. Parallelization: Uses ThreadPoolExecutor to process all face groups simultaneously.

Setup

Install Dependencies:

pip install -r requirements.txt

Note: opencv-contrib-python is recommended for full feature access.

Folder Structure: Create a data folder in the project root. Inside, structure it as follows:

/data
    /face1
        image1.jpg
        image2.jpg
    /signface1
        closeup_sign.jpg
    /face2
        ...
    ...

Run the System:

python sign_pipeline.py

Customization

Integrating Real AI Models

To replace the mock logic with a real model (e.g., YOLOv8-seg or Mask R-CNN):

Open sign_pipeline.py.

Locate the MockSignDetector class.

Modify detect_segmentation to run your model inference and return binary numpy masks.

Performance Tuning

Threads: Adjust max_workers in PipelineProcessor.run() based on your CPU cores.

Stitching: If stitching is too slow, resize input images in load_images_from_folder before returning them.