From Table \ref{995837}, we can see improvement of 6% in mAP by comparing Model A with Model B . In this project, all tobacco advertisements, regardless of brand or message, have been grouped as a single class. Both Model A and Model B were trained and tested with MD17 dataset. The only difference between model A and model B are the anchor scales. Model C was adopted from Model B by continuing training and testing using the PANIC18 dataset. PANIC18 was also derived from Google Street View but annotated and manipulated according to the procedure described in this project. We found that PANIC18 has more variety in terms of tobacco promotions and products, including electronic cigarettes, while MD17 focused almost exclusively on traditional cigarette brands like Marlboro and Newport. Since PANIC18 is a smaller dataset, Model C fell short of learning the features of less common products.
From Figure \ref{640957}, the model detected all four traditional cigarette brands signs in one image with high probabilities. Numbers next to bounding boxes are the probabilities of such objects being tobacco advertisements. Similarly, the traditional "Newport" and "Marlboro" brand signs were detected as well. Such detections are typical examples that were found from our model output. We used a classification probability of 0.7 for the model.
From Figure \ref{360814}, we can see that our model incorrectly detected street signs, blurry boxes, and traffic cones which could be impacted by the sign and similarities show in the bottom row. They share features like color combination, shapes, and sizes. Other false detection examples include trash bins, windows, and glass doors.