Towards accurate billboard detection: an ablation and benchmarking study of deep learning models

With the increase in multimedia videos, the applications of videos and images to leverage targeted advertisements are becoming significantly diverse in reaching specific audiences. Targeted advertisements are achieved by detecting and replacing an existing advertisement within an image frame with a new advert. In this work, we classify image frames as advert or no-advert. We conducted a benchmarking study using two datasets: the AdShop dataset and the ALOS dataset. Our evaluation indicates that the AdShop dataset is suboptimal for billboard detection. Conversely, the ALOS dataset, specifically curated for billboard identification, demonstrates superior accuracy and reliability. We also performed an ablation study by reducing the filter size at different layers to estimate the classification performance. The results suggest that the proposed ModifiedVGGNet achieves the highest performance with a training accuracy of 97.2% and a test accuracy of 97.0%. This was achieved by reducing 10% filter in conv1 layer of block 1.