Automatic Signboard Detection and Localization in Densely Populated Developing Cities

Most city establishments of developing cities are unlabeled because of the necessity of manual annotation. Hence location and trajectory services remain under utilized in such cities. Accurate signboard detection and localization in natural scene images is the foremost task for accurate information retrieval from such city streets. We develop an automated signboard detection system suitable for such cities using Faster R-CNN based localization by incorporating two specialized pretraining methods and a run time efficient hyperparameter value selection algorithm. We have taken an incremental approach in reaching our final proposed model through detailed evaluation and comparison with baselines using our constructed SVSO signboard dataset containing signboard natural scene images of six developing countries. Our proposed method can detect signboards accurately, even though images contain multiple signboards with diverse shapes and colours in a noisy background. Our proposed model achieves 0.91 mAP score on validation set and 0.90 mAP score on an independent test set.