We use synthetic dataset of real-world rainy images for training and testing. Rain100L contains 200 training image pairs and 100 testing image pairs, and the artificially added rain streaks have only a single mode. Rain100H is more complicated and more challenging. It contains 1,254 training image pairs and 100 image pairs for testing, and the artificially added rain streaks have five modes. Real-world rainy images, Provided by W. Yang et al are tested directly.

Most of the Chinese characters are composed of radicals and semantic components. Using Nonnegative Matrix Factorization, we managed to train the computer to recognize and disassemble the characters into various components and reassemble them again. This technique can be useful for computers to understand the balance of various components in a character, which could be complicated in some case. It is possible to design Chinese font easily when computers can generate and assemble the components, reducing the need to actually designing more than 8,000 characters.

The relationship between humans or groups can be complicated. By surveilling their interactions, it is possible to map out the degree of closeness and their relationship. In this image, the characters of a movie are numbered, while the lines between them shows the occurrence which the characters interact with each other. This research has the potential to develop more advanced recognition method, probably elaborating the relationship by understanding facial features or body gestures.

To increase the efficiency of stores management using robots, Brain Corp and Walmart are scaling up from an initial 360 robotic floor cleaner trial to add 1,500 more robots. This method can be achieved by identifying details and dimension of a space to map out the pathways and storage area, allowing robots to do regular cleaning and restocking.

The density of convenient stores in Taiwan is high, and it is easy to find them in towns and residence for their service. However, the blinds would need guidance to visit the stores. By choosing 14 types of signboards of some popular stores, we trains the machine to identify them in maps obtained from Google Map. By identifying where the stores are located, it will facilitates the blinds to visit the stores by knowing their location and distance without too relying on others.

Precise vertebral segmentation provides the basis for spinal image analyses and interventions, such as vertebral compression fracture detection and other abnormalities. We proposed an iterative vertebrae instance segmentation model, which has good generalization ability for segmenting all types of vertebrae, including cervical, thoracic, and lumbar vertebrae. Our model not only used 17% less memory, but also achieved better performance on vertebrae segmentation compared to existing methods. Additionally, our model can provide additional output for accurate anatomical prediction under the same amount of memory.