Logo Detection in Images and Videos

Objective

Service to detect brand logos from an unknown dataset of images and videos.

Compiled a diverse training dataset using both public and internal data sources.
Converted the labeling to a binary classification: ‘Has Logo’ or ‘No Logo’.
Trained a YOLOv7 model on over 40,000 images.
Evaluated model performance using the mean Average Precision (mAP) metric; however, initial results showed that detected logos had low confidence scores.
Addressed the high number of false positives by integrating a template matching process using Canny edge detection to dynamically adjust the threshold and match the detected logos with brand logo templates provided.
This process significantly reduced false positives by refining the final set of detected logos.

Despite the low confidence scores in detections, the overall pipeline achieved a high accuracy rate of 93% in detecting logos from an unseen dataset.