A5023601808- Advanced Vision and Imaging
- Advanced Image Processing Techniques
- Optical measurement and interference techniques
- Image Processing Techniques and Applications
- Image Enhancement Techniques
- Remote Sensing and LiDAR Applications
- Robotics and Sensor-Based Localization
- Industrial Vision Systems and Defect Detection
- Surface Roughness and Optical Measurements
- Infrastructure Maintenance and Monitoring
- Advanced Neural Network Applications
- Computer Graphics and Visualization Techniques
- Advanced Measurement and Metrology Techniques
- Digital Holography and Microscopy
- Fluid Dynamics and Turbulent Flows
- Medical Image Segmentation Techniques
- Handwritten Text Recognition Techniques
- 3D Shape Modeling and Analysis
- Generative Adversarial Networks and Image Synthesis
- Image and Signal Denoising Methods
- Manufacturing Process and Optimization
- Advanced Image and Video Retrieval Techniques
- Electronic Packaging and Soldering Technologies
- Video Coding and Compression Technologies
- Optical Coherence Tomography Applications
Chinese University of Hong Kong
2011-2020
University of Hong Kong
2007-2008
FlowNet2 [14], the state-of-the-art convolutional neural network (CNN) for optical flow estimation, requires over 160M parameters to achieve accurate estimation. In this paper we present an alternative that attains performance on par with challenging Sintel final pass and KITTI benchmarks, while being 30 times smaller in model size 1.36 faster running speed. This is made possible by drilling down architectural details might have been missed current frameworks: (1) We a more effective...
This paper reviews the 2nd NTIRE challenge on single image super-resolution (restoration of rich details in a low resolution image) with focus proposed solutions and results. The had 4 tracks. Track 1 employed standard bicubic downscaling setup, while Tracks 2, 3 realistic unknown downgrading operators simulating camera acquisition pipeline. were learnable through provided pairs high train images. tracks 145, 114, 101, 113 registered participants, resp., 31 teams competed final testing...
Over four decades, the majority addresses problem of optical flow estimation using variational methods. With advance machine learning, some recent works have attempted to address convolutional neural network (CNN) and showed promising results. FlowNet2 [1] , state-of-the-art CNN, requires over 160M parameters achieve accurate estimation. Our LiteFlowNet2 outperforms on Sintel KITTI benchmarks, while being 25.3 times smaller in model size 3.1 faster running speed. is built foundation laid by...
We study the problem of distilling knowledge from a large deep teacher network to much smaller student for task road marking segmentation. In this work, we explore novel distillation (KD) approach that can transfer `knowledge' on scene structure more effectively model. Our method is known as Inter-Region Affinity KD (IntRA-KD). It decomposes given image into different regions and represents each region node in graph. An inter-region affinity graph then formed by establishing pairwise...
Unsupervised methods have showed promising results on monocular depth estimation. However, the training data must be captured in scenes without moving objects. To push envelope of accuracy, recent tend to increase their model parameters. In this paper, an unsupervised learning framework is proposed jointly predict and complete 3D motion including motions objects camera. (1) Recurrent modulation units are used adaptively iteratively fuse encoder decoder features. This improves single-image...
FlowNet2, the state-of-the-art convolutional neural network (CNN) for optical flow estimation, requires over 160M parameters to achieve accurate estimation. In this paper we present an alternative that outperforms FlowNet2 on challenging Sintel final pass and KITTI benchmarks, while being 30 times smaller in model size 1.36 faster running speed. This is made possible by drilling down architectural details might have been missed current frameworks: (1) We a more effective inference approach...
Over four decades, the majority addresses problem of optical flow estimation using variational methods. With advance machine learning, some recent works have attempted to address convolutional neural network (CNN) and showed promising results. FlowNet2, state-of-the-art CNN, requires over 160M parameters achieve accurate estimation. Our LiteFlowNet2 outperforms FlowNet2 on Sintel KITTI benchmarks, while being 25.3 times smaller in model size 3.1 faster running speed. is built foundation laid...
A two-step phase shift profilometry method (2-step PSP) with prefiltering and postfiltering stages is proposed to reconstruct the 3-D profile of solder paste. Two sinusoidal patterns which are pi-out-of-phase used in reconstruction. The new uses only two fringe rather than four as four-step (4-step PSP). In Fourier transform (FTP), a bandpass filter required extract fundamental spectrum from background higher order harmonics due camera noise imperfectness pattern projector. By using...
Depth maps from low-cost RGB-D system are generally noisy and not accurate enough. Holes often exist in the depth maps. Bilateral filter is commonly utilized to perform enhancement. However, it requires high computational time. Its texture transferring property also makes those boundaries between textured homogeneous regions filtered map far satisfactory. In this paper, we present a method raw using guided filtering two-stage framework. Our only has faster time than bilateral but avoids...
We address the problem of recovering camera motion from video data, which does not require establishment feature correspondences or computation optical flows but normal directly. have designed an imaging system that has a wide field view by fixating number cameras together to form approximate spherical eye. With substantially widened visual field, we discover estimating directions translation and rotation components separately are possible particularly efficient. In addition, inherent...
Low-cost RGB-D imaging system such as Kinect is widely utilized for dense 3D reconstruction. However, generally suffers from two main problems. The spatial resolution of the depth image low. often contains numerous holes where no measurements are available. This can be due to bad infra-red reflectance properties some objects in scene. Since color higher than that image, this paper introduces a new method enhance images captured by moving using cues induced optical flow. We not only fill raw...
Phase profilometry is an effective and efficient way to reconstruct 3D profile of objects through optical method. It requires a pattern generator project sinusoidal onto objects. The deformed then captured by camera. However, noise from camera, distortion lens, imperfectness projector cause trouble the reconstruction. necessary filter images before applying phase profilometry. In this paper, 4-step shift (PSP) used solder paste. A 2D median ID non-linear filters (pre-filtering) are proposed...
We study the problem of distilling knowledge from a large deep teacher network to much smaller student for task road marking segmentation. In this work, we explore novel distillation (KD) approach that can transfer 'knowledge' on scene structure more effectively model. Our method is known as Inter-Region Affinity KD (IntRA-KD). It decomposes given image into different regions and represents each region node in graph. An inter-region affinity graph then formed by establishing pairwise...
Existing unconditional generative models mainly focus on modeling general objects, such as faces and indoor scenes. Fashion textures, another important type of visual elements around us, have not been extensively studied. In this work, we propose an effective model for fashion textures also comprehensively investigate the key components involved: internal representation, latent space sampling generator architecture. We use Gram matrix a suitable representation realistic further design two...
In this paper, we address the problem of dense depth map generation from two successive image frames in a video. We first recover camera motion observable normal flow pattern using our previously proposed apparent constraints. Once is estimated, sparse data can be directly recovered pattern. utilize hierarchical approach to generate an initial data. This further enhanced through refinement associated optical field variational framework. Experimental results show that method provide...
We address the problem of motion recovery and depth map reconstruction from a sequence stereo images. The proposed method utilizes normal flows which are directly observable in image sequence. No explicit correspondence establishment is required. first determine direction translation full rotation for each cameras system using an iterative algorithm. Then we recover magnitude geometrical method. Depth metric scale reconstructed once parameters estimated. Experimental results on synthetic...
We present a direct method for estimating general camera motion from rectified stereo image sequence using merely normal flows. require neither point-to-point feature correspondences nor special assumption about the imaged scene. It has closed-form expression, requiring no iterative computation. Experimental results on both synthetic and real data are provided to demonstrate performance of proposed method.