Transfer Learning for Object Detection using State-of-the-Art Deep Neural Networks

Transfer learning through the use of synthetic images and pretrained convolutional neural networks offers a promising approach to improve the object detection performance of deep neural networks. In this paper, we explore different strategies to generate synthetic datasets and subsequently improve them to achieve better object detection accuracy (mAP) when trained with state-of-the-art deep neural networks, focusing on detection of packed food products in a refrigerator. We develop novel techniques like dynamic stacking, pseudo random placement, variable object pose, distractor noise etc. which not only aid in diversifying the synthetic data but also help in improving the overall object detection mAP by more than 40%. The synthetic images, generated using Blender-Python API, are clustered in a variety of configurations to cater to the diversity of real scenes. These datasets are then utilized to train TensorFlow implementations of state-of-the-art deep neural networks like Faster-RCNN, R-FCN, and SSD and their performance is tested on real scenes. The object detection performance of various deep CNN architectures is also studied, with Faster-RCNN proving to be the most suitable choice, achieving the highest mAP of 70.67.