A deep framework for automatic annotation with application to retail warehouses

The paper presents a novel deep learning framework for automatic annotation and segmentation of densely cluttered objects in a warehouse application use-case as specified by the Amazon Robotics Challenge (ARC) 2017. This framework addresses two challenges of the competition: (1) reducing the amount of manual labour involved in generating a large number of annotated data that could be used for training a deep network and, (2) achieving good segmentation accuracy in a very limited amount of training time ( 30 minutes). These two problems are solved by proposing a deep architecture comprising of Residual Network and Feature Pyramidal based convolutional neural network that helps to retain primitive features along with higher level features obtained from each successive layer. In addition, a framework is proposed using this network to automatically generate a large annotated dataset having different degrees of clutters to carry out multi-class semantic segmentation after training with this machine generated dataset. The proposed framework is shown to provide better segmentation accuracy with lesser training time as compared to the existing state-of-the-art architectures such as PSPNet and Mask R-CNN. The overall working of the proposed architecture is explained by creating a new dataset from the objects specified by the ARC competition. An extensive experiment is also performed using the MIT-Princeton database [22]. Our TCS-ARCDataset [13]is made available online for the convenience of readers.