Deep learning for combo object detection

Convolutional neural networks (CNNs) have become the most vigorous technique for a variety of different tasks in computer vision, due to their proficiency in automatically learning high-level visual representations for images. In this paper, we investigate the effect of deep neural networks on the accuracy in combo object detection setting. The insufficiency of labeled data, coupled with the uncertainty of spacial distribution and dynamic changes in luminance, creates situations where combo object detection is far more challenging. Using transfer learning, we present a system for combo object detection based on a deep CNN called ComboNN. The proposed ComboNN is pre-trained on a huge auxiliary dataset ImageNet and fine-tuned on our small dataset. The use of data augmentation and regularization technique significantly reduces overfitting and improves the robustness of the ComboNN. Experimental results demonstrate that our system is capable of making reliable prediction on combo object detection in the real-world images, and achieves much better accuracy than the state-of-the-art CNNs.