joaosousapt/galo_de_barcelos

Galo de Barcelos Detection

Overview

The main goal of this side project is to perform robust object detection of the "Galo de Barcelos" (The Rooster of Barcelos), one of the most iconic symbols of Portugal. Why this project? Unlike mass-produced industrial items, each "Galo de Barcelos" is unique. They are hand-painted with a huge diversity of patterns, colors, shapes, and artistic details. This high variability presents a unique computer vision challenge:

• Artistic Generalization: Testing how far a compact object detector can generalize to handmade, non-standardized objects.
• Complex Patterns: Dealing with intricate "hearts" and floral motifs that vary from artist to artist.
• Compact Efficiency: Optimizing for a model that stays "lightweight" without losing the ability to recognize these rich details.

This is a personal side project and it's still in a very early phase.

The dataset is currently very small. Because of the limited data, i'm mainly experimenting with frozen backbone fine-tuning, this usually gives more stable results and much less overfitting when you don't have many examples. The results are not perfect yet (some missed detections, false positives, inconsistent confidence scores, sensitivity to lighting / angle / background), this is completely expected with such a small dataset.

https://github.com/joaosousa1/Browser-ONNX-Image-Inference