Barbados Land Detection and OCR Project - Solution Summary

Code: https://github.com/josephgitau/Barbados-Lands-and-Surveys-Plot-Automation-Challenge/tree/main

Platform

• Dual-notebook workflow executed on Colab GPUs (L4 for training, A100 High Ram for inference).
• Total pipeline execution time: ~6.5 hours.

Data Aspects

• Source: cadastral survey map datasets.
• Format: JPEG/PNG images with COCO-style annotation JSON (results.json).
• Annotation: Used Label Studio for polygon mask annotation on 658 images.

Preprocessing and Post-processing

• Image Preprocessing: Resizing to 1024x1024, encoder-specific normalization (Albumentations), and data augmentation (flips, rotations, color perturbations).
• Post-processing: Polygons cleaned with steps like removing self-intersections, simplifying boundaries (RDP = 0.0025), and smoothing near-straight lines.

Model Architecture

• Segmentation: UNet++ with an EfficientNet-B7 encoder.
• OCR: Large Vision-Language Model Qwen3-VL-30B-A3B-Instruct-FP8 (used via vLLM runtime).
• Training: PyTorch Lightning, Adam optimizer, Mixed Precision (AMP 16).
• Loss Function: Combination of Boundary Loss, Focal Loss, and optional Dice/BCE.

Insight

• The choice of the large, pretrained Qwen3-VL-30B for OCR was intentional to leverage its strong zero-shot generalization and avoid the risks of overfitting and memorization associated with Supervised Fine-Tuning (SFT) on noisy data.
• The segmentation and OCR models combined achieved a Private Leaderboard Score of 0.970242006.
• The dual-pipeline delivers reproducible outputs (barbados_final.csv) combining cleaned polygon geometries and OCR-extracted text.

Final Comments

The dataset was flawed from the start (you can check the data cleanup notebook for all stats), communication was an issue and rules were flawed in this competiotion. Despite these hurdles, the competition was an invaluable learning opportunity. The skills we have attained are priceless, and that matter more.

See you in the next one.