Yasunori Ishii

Analysis of photometric factors based on photometric linearization.

We propose a method for analyzing photometric factors, such as diffuse reflection, specular reflection, attached shadow, and cast shadow. For analyzing real images, we utilize the photometric linearization method, which was originally proposed for image synthesis. First, we show that each pixel can be photometrically classified by a simple comparison of the pixel intensity. Our classification algorithm requires neither 3D shape information nor color information of the scene. Then, we show that the accuracy of the photometric linearization can be improved by introducing a new classification-based criterion to the linearization process. Experimental results show that photometric factors can be correctly classified without any special devices. A further experiment shows that the proposed method is effective for photometric stereo.

Classification of photometric factors based on photometric linearization

We propose a new method for classification of photometric factors, such as diffuse reflection, specular reflection, attached shadow, and cast shadow. For analyzing real images, we utilize the photometric linearization method which was originally proposed for image synthesis. First, we show that each pixel can be photometrically classified by the simple comparison of the pixel intensity. Our classification algorithm requires neither 3D shape information nor color information of the scene. Then, we show that the accuracy of the photometric linearization can be improved by introducing a new classification-based criterion to the linearization process. Experimental results show that photometric factors can be correctly classified without any special device.

Multi-staged deep learning with created coarse and appended fine categories

This paper proposes a new learning method for Deep Learning based on the concept of a Coarse-to-Fine approach. The Coarse-to-Fine classification improves Deep Learning performance, but it increases network size and presents the problem of close dependence on the accuracy of coarse classification. We tried to avoid this problem by adopting the concept of Curriculum Learning and succeeded in improving the accuracy of Deep Learning. This technique uses learning that employs a single closed image dataset several times in the same network except for the last layer. In this process, coarse labels are given to the images during the pre-training stages and fine labels are given to the same images at the fine-tuning stage. This coarse category pre-training method makes it possible to obtain those features that commonly exist in multiple fine categories. To demonstrate the advantage of this technique, several patterns of a dataset in the quantity of several tens of classes and a single dataset of 100 classes were produced using the ImageNet dataset and compared with the previous technique. The results showed a 5.7% improvement of TOP1 accuracy, with the best case confirmed in the 100-class dataset.

Profile face detection using block difference feature for automatic image annotation

We propose a new profile face detection system for automatic image annotation. For personal images, profile face detection is required because faces turn in different directions. To reduce the numerous false detections seen in previous approaches, we propose a block difference feature. The experiments we describe show the successful outcome of this strategy.