yiyuezhang

Professor Kuo Gave a Talk on Deep Learning at Ming Hsieh Institute

The Ming Hsieh Institute has launched an MHI Emerging Trends Series. MCL Director, Professor C.-C. Jay Kuo, was the first speaker in this series. Professor Kuo gave his talk on deep learning on April 10 (Monday), 2017.

There is a resurging interest in developing a neural-network-based solution to supervised machine learning in the last 5 years. However, little theoretical work was reported in this area. In his talk, Professor Kuo attempted to provide some theoretical foundation to the working principle of the convolutional neural network (CNN) from a signal processing viewpoint. First, he introduced the RECOS transform as a basic building block for CNNs. The term “RECOS” is an acronym for “REctified-COrrelations on a Sphere”. It consists of two main concepts: data clustering on a sphere and rectification. Then, a CNN is interpreted as a network that implements the guided multi-layer RECOS transform. Along this line, he also compared the traditional single-layer and modern multi-layer signal analysis approaches. Furthermore, he discussed how guidance can be provided by data labels through backpropagation in the training with an attempt to offer a smooth transition from weakly to heavily supervised learning. Finally, he pointed out several future research directions at the end.

There were about 80 people attending Professor Kuo’s seminar. Many questions were asked after his talk. Professor Kuo said that he enjoyed the interaction with the audience very much and it demonstrated the strong interest of the audience on this topic.

MCL members, Chi-Hao (Eddy) Wu and Siyang Li presented their papers at Winter Conference on Applications of Computer Vision (WACV) 2017, Santa Rosa, CA, USA

The title of Eddy’s paper is “Boosted Convolutional Neural Networks (BCNN) for Pedestrian Detection”, with Weihao Gan, De Lan and C.-C. Jay Kuo as the co-authors. Here is a brief summary:

“In this work, a boosted convolutional neural network (BCNN) system is proposed to enhance the pedestrian detection performance. Being inspired by the classic boosting idea, we develop a weighted loss function that emphasizes challenging samples in training a convolutional neural network (CNN). Two types of samples are considered challenging: 1) samples with detection scores falling in the decision boundary, and 2) temporally associated samples with inconsistent scores. A weighting scheme is designed for each of them. Finally, we train a boosted fusion layer to benefit from the integration of these two weighting schemes. We use the Fast-RCNN as the baseline, and test the corresponding BCNN on the Caltech pedestrian dataset in the experiment, and show a significant performance gain of the BCNN over its baseline.” 

Siyang’s paper is entitled “Box Refinement: Object Proposal Enhancement and Pruning”, co-authored with Heming Zhang, Junting Zhang, Yuzhuo Ren and C.-C. Jay Kuo. The summary goes as followed:

“Object proposal generation has been an important preprocessing step for object detectors in general and the convolutional neural network (CNN) detectors in particular. Recently, people start to use the CNN to generate object proposals but most of these methods suffer from the localization bias problem, like other objectness-based methods. Since contours offer a powerful cue for accurate localization, we propose a box refinement method by searching for the optimal contour for each initial bounding box that minimizes the contour cost. [...]

Fashion fingerprint is a compact feature vector for fashion items that can be used for tasks such as recognition, clustering, and retrieval of similar items. It is equally useful for both online fashion retailers as well as for physical apparel stores (with or without their online extensions). A related problem is understanding the apparel preferences of an individual from the dresses that they wear while visiting the physical store. One of the challenges of fashion study different from others is lack of enough accurate annotation. Available datasets have either limited number of images or very noisy annotation.

Currently we have successfully trained a fashion item localization model based on SSD[1]. The model is able to localize upper clothes, bottom clothes and one-pieces and has been tested on the Clothing Parsing dataset[2]. It achieves an F-score of 0.887 on upper clothes localization. For other clothing items, errors occur because the model may focus on too local regions and thus gets confused between skirt and dress. Prior location of human body will be incorporated in our model to solve this problem.

In the future we will further refine our localization model and also work on two directions. One is to recognize the garments based on our localization. The other one is to automatically label more images to enlarge the size of datasets.

[1] Liu, Wei, et al. “SSD: Single shot multibox detector.” European Conference on Computer Vision. Springer International Publishing, 2016.
[2] Liang, Xiaodan, et al. “Deep human parsing with active template regression.” IEEE transactions on pattern analysis and machine intelligence 37.12 (2015): 2402-2414.

Automatic image segmentation has always been an important topic in medical imaging. Many medical applications, such as delineating heart structures, rely heavily on the accurate segmentation results. Nowadays, manual segmentation is still required in many applications. Manual segmentation is not only time-consuming and tedious but also prone to human error. One of MCL members, Ruiyuan Lin, is working on this research topic.

Many methods have been proposed to automate the segmentation process, ranging from region growing and active contour models to multi-atlas segmentation. In our research work, we focus on the convolutional neural networks (CNN) based segmentation method. We attempted several segmentation networks such as fully convolutional networks (FCN) and residual networks, compared their performance with other methods, and analyzed the strengths and problems of the networks. We are planning to further explore the use of CNN on more complicated medical images such as cross-domain images.

Image credit: both images are modified from the MRI images in the Left Atrium Segmentation Challenge dataset:
Tobon-Gomez C, Geers AJ, Peters, J, Weese J, Pinto K, Karim R, Ammar M, Daoudi A, Margeta J, Sandoval Z, Stender B, Zheng Y, Zuluaga, MA, Betancur J, Ayache N, Chikh MA, Dillenseger J-L, Kelm BM, Mahmoudi S, Ourselin S, Schlaefer A, Schaeffter T, Razavi R, Rhode KS. Benchmark for Algorithms Segmenting the Left Atrium From 3D CT and MRI Datasets. IEEE Transactions on Medical Imaging, 34(7):1460–1473, 2015.

Advanced driver assistance systems (ADAS) have attracted more and more attention nowadays, where various IT technologies are introduced to vehicles to enhance driving safety and automation.

MCL members, Junting Zhang and Yuhang Song, together with MediaTek Inc. have started a collaborative research project on ADAS-oriented deep learning technologies since January 2016. Single-image-based traffic scene segmentation and road detection have been studied extensively throughout 2016. We adapted the state-of-the-art general-purpose CNN architectures to urban scene semantic segmentation task, overcoming the cross-domain issue. On the other hand, computational and memory efficiency have always been our major concerns, we were also devoted to simplify the network structure and reduce redundant computation.

In 2017, we will explore the deep learning technologies for video processing. Although there are many interesting results in semantic urban scene understanding based on the CNN technology, semantic video understanding is still a challenging problem. We will try to find a semantic video understanding method that outperforms the single-image-based algorithms. To address this type of problems, we will exploit the temporal information.