DiffNet: A learning to compare deep network for product recognition

The paper focuses on the identification of different objects in a pair of images taken from the same environment, which is challenging and has wide application. We propose a single deep convolutional neural network termed as DiffNet to solve this problem. DiffNet takes a pair of images as the input and directly regresses the bounding boxes of different objects. To train DiffNet, we only need to label the different objects, rather than all objects in input images, which significantly reduces human labeling efforts. Experiments are performed on an image dataset collected from unmanned containers. DiffNet obtains a very high product detection accuracy of 95.56% mAP at the speed of 143 fps measured on an NVIDIA TITAN Xp GPU.