Temporal action localization is an important yet challenging problem. Given a long, untrimmed video consisting of multiple action instances and complex background contents, we need not only to recognize their action categories, but also to localize the start time and end time of each instance. Many state-of-the-art systems use segment-level classifiers to select and rank proposal segments of pre-determined boundaries. However, a desirable model should move beyond segment-level and make dense predictions at a fine granularity in time to determine precise temporal boundaries. To this end, we design a novel Convolutional-De-Convolutional (CDC) network that places CDC filters on top of 3D ConvNets, which have been shown to be effective for abstracting action semantics but reduce the temporal length of the input data. The proposed CDC filter performs the required temporal upsampling and spatial downsampling operations simultaneously to predict actions at the frame-level granularity. It is unique in jointly modeling action semantics in space-time and fine-grained temporal dynamics. We train the CDC network in an end-to-end manner efficiently. Our model not only achieves superior performance in detecting actions in every frame, but also significantly boosts the precision of localizing temporal boundaries. Finally, the CDC network demonstrates a very high efficiency with the ability to process 500 frames per second on a single GPU server. Source code and trained models are available online at https://bitbucket.org/columbiadvmm/cdc.

This paper proposes a noise reduction method for screen content coding using HEVC. The proposed method focuses on that the histograms of pixel values for screen content have distinct and sparse peaks respectively. This property is often degraded by the coding distortion, which arises some small peaks around the original peak. This paper presents a method for removing the small peaks derived from the noise, and its implementation to HEVC as an extension of SAO. Experimental evaluation shows that the proposed noise reduction method significantly improves the visual quality of decoded screen content with bit-rate reduction.

This paper describes a hardware implementation of HEVC (High Efficiency Video Coding) video encoder for exploring performance of HEVC in real-time applications. By optimizing hardware architecture and developing fast mode decision algorithms, our FPGA-based prototype can encode 1080p, 10-bit video at 60fps in real-time. Both objective and subjective evaluations clearly demonstrate that the developed HEVC encoder achieves significant coding gain relative to professional-use AVC/H.264 encoder available on the market. This paper also refers to the development of a real-time HEVC encoder for much higher resolution video such as SHV (Super Hi-Vision, 8K) using the HD encoder in parallel.

This paper presents a method for reducing the complexity of adaptive filtering for image restoration adopted in High Efficiency Video Coding (HEVC), called Adaptive Loop Filter (ALF). Although ALF can greatly improve the PSNR of decoded frames, its impact is not uniform across an entire frame, but depends on the local texture. Based on this fact, the proposed method first predicts the effectiveness of ALF for each pixel by score calculation, and skips all ALF processes for the pixels whose scores are less than a threshold. This threshold is adaptively determined for each frame in the encoding process, and is transmitted to a decoder. Experimental evaluation clearly demonstrates that the proposed method achieves significant reduction of the computation time of ALF without noticeable encoding performance loss.

This paper proposes a novel method for aligning two signals using the information contained in the overlapped band. In particular the proposed method aligns two signals by compensating both time-delay and phase-offset in the second signal using the estimated gradient of phase difference in the overlapped band. Compared with other conventional methods, this method can align two signals without requiring a pilot tone or additional hardware. The proposed method was experimentally validated using RF pulses.

This paper presents a palmprint recognition algorithm using principal component analysis (PCA) of phase information in 2D (two-dimensional) discrete Fourier transforms (DFTs) of palmprint images. To achieve highly robust palmprint recognition, the proposed algorithm (i) limits the frequency bandwidth, and (ii) averages phase spectra using multiple palmprint images captured from the same hand at an enrollment stage. Through a set of experiments, we demonstrate that the proposed method can significantly reduce computational cost without sacrificing recognition performance compared with our previous work using phase-only correlation (POC) - an image matching technique using the phase components in 2D DFTs of given images. Also, the resulting performance is much higher than those of conventional palmprint recognition algorithms which apply PCA to palmprint images, or phase spectra directly.

This paper proposes a three-dimensional (3D) body scanning system that uses passive stereo vision with a robot arm. So far, the reported 3D body scanning systems employ active 3D measurement methods. However, active methods use structured illumination or laser scanning, which is not desirable in many systems applied to human. A major problem of using passive stereo vision for 3D measurement is its low accuracy. In addition, multiple stereo images captured from different viewpoints are necessary to cover the whole body at an appropriate distance. Addressing these problems, we have newly developed an eye-in-hand system based on passive stereo vision, where a phase-based image matching technique is employed for sub-pixel disparity estimation. Through a set of experiments, we demonstrate that the proposed system can capture 3D shape of human body with high quality.

This paper presents an efficient algorithm for iris recognition using phase-based image matching--an image matching technique using phase components in 2D Discrete Fourier Transforms (DFTs) of given images. Experimental evaluation using CASIA iris image databases (versions 1.0 and 2.0) and Iris Challenge Evaluation (ICE) 2005 database clearly demonstrates that the use of phase components of iris images makes possible to achieve highly accurate iris recognition with a simple matching algorithm. This paper also discusses major implementation issues of our algorithm. In order to reduce the size of iris data and to prevent the visibility of iris images, we introduce the idea of 2D Fourier Phase Code (FPC) for representing iris information. The 2D FPC is particularly useful for implementing compact iris recognition devices using state-of-the-art Digital Signal Processing (DSP) technology.

This paper presents an implementation of iris recognition algorithm using phase-based image matching-an image matching technique using phase components in 2D discrete Fourier transforms (DFTs) of given images. Our experimental observation clearly shows that the use of phase components of iris images makes possible to achieve highly accurate iris recognition even for low-quality iris images. In this paper, we consider the problem of designing a compact phase-based iris recognition algorithm especially suitable for hardware implementation. We also present prototype implementation of an iris recognition system based on the proposed algorithm. The prototype system fully utilizes state-of-the-art DSP (digital signal processor) technology to achieve real-time iris recognition capability within a compact hardware module.

A major approach for iris recognition today is to generate feature vectors corresponding to individual iris images and to perform iris matching based on some distance metrics. One of the difficult problems in feature-based iris recognition is that the matching performance is significantly influenced by many parameters in feature extraction process, which may vary depending on environmental factors of image acquisition. This paper presents an efficient algorithm for iris recognition using phase-based image matching. The use of phase components in 2D (two-dimensional) discrete Fourier transforms of iris images makes possible to achieve highly robust iris recognition in a unified fashion with a simple matching algorithm. Experimental evaluation using an iris image database clearly demonstrates an efficient matching performance of the proposed algorithm.