Application of Transfer Learning for Object Detection on Manually Collected Data

This paper investigates the usage of pre-trained deep learning neural networks for object detection on a manually collected dataset for real-life indoor objects. Availability of object-specific datasets is a great challenge and the unavoidable task of collecting, processing and annotating ground truth data is laborious and time-consuming. In this paper, two famous models (AlexNet and Vgg16) have been evaluated as feature extractors in a Faster R-CNN network. Network models have been trained end-to-end on the collected dataset. The study highlights the poor performance of state of art systems when dealing with small size objects. Modifying the detector design by redesigning systems’ anchor boxes might help to tackle this problem. Detector results on the proposed dataset have been collected and compared. In addition, limitations and future work have been discussed.