CVPR 2020: Watch MediaTek's Presentations Online

The CVPR (IEEE Conference on Computer Vision and Pattern Recognition) conference is the premier annual AI event attended by the world’s foremost technology companies in the fields of Computer Vision, Machine Learning and AI, as well as associated academics, researchers, and students. CVPR is taking place between June 14-19th 2020, with presentations available to watch online.

Click here for the full programs of events >

In 2020, CVPR has accepted several papers authored by MediaTek teams:

• Unified Dynamic Convolutional Network for Super-Resolution with Variational Degradations
• Learning to See Through Obstructions
• Single-Image HDR Reconstruction by Learning to Reverse the Camera Pipeline

Along with the CVPR conference, MediaTek is also sponsoring associated workshops at NTIRE’20 (New Trends in Image Restoration and Enhancement). Here, MediaTek teams will present an accepted paper regarding advanced image restoration on mobile devices (Click here for more information):

• Deploying Image Deblurring across Mobile Devices: A Perspective of Quality and Latency
  

MediaTek is also sponsoring the LPCV’20 (Low Power Computer Vision) and CLIC’20 (Challenge on Learned Image Compression) workshop to push the boundaries of Edge AI. If you’re an AI expert looking for new opportunities, MediaTek has careers available.

The following is a brief introduction of each paper:

CVPR’20: Unified Dynamic Convolutional Network for Super-Resolution with Variational Degradation

Online live presentation and Q&A times for this paper:

6/18 15:00-17:00 (PDT, Seattle Time) / 6/18 3:00-5:00 (PDT, Seattle Time)

Overview:

The goal of single image super-resolution (SISR) is to reconstruct a High Resolution (HR) image from a Low-Resolution (LR) image. The recent advances in computer vision – such as Deep Convolutional Neural Networks (CNNs) - has achieved remarkable results in retaining and improving image quality when performing SISR tasks. Specifically, MediaTek teams have used Dynamic Convolution to fix variational degradations.

In this paper, MediaTek has proposed a unified framework UDVD (Unified Dynamic Convolutional Network for Variational Degradations) that uses dynamic convolution to handle variational cross-image and intra-image degradation that includes noise, blur and downsampling for Super Resolution applications. The paper demonstrates a comprehensive analysis of the performance impact of dynamic convolution and our extensive experiments show that the proposed UDVD achieves more favorable performance than state-of-the-art methods on both synthetic and real images.

CVPR’20: Learning to See Through Obstructions

Online live presentation and Q&A times for this paper:

6/18 17:00-19:00 (PDT, Seattle Time) / 6/18 5:00-7:00 (PDT, Seattle Time)

Overview:

MediaTek and collaborative researchers present a learning-based approach for removing unwanted obstructions, such as window reflections, fence occlusions or raindrops, from a short sequence of images captured by a moving camera. Our method leverages the motion differences between the background and the obstructing elements to recover both layers. Specifically, we alternate between estimating dense optical flow fields of the two layers and reconstructing each layer from the flow-warped images via a deep convolutional neural network. The learning-based layer reconstruction allows us to accommodate potential errors in the flow estimation and brittle assumptions such as brightness consistency. We show that training on synthetically generated data transfers well to real images. Our results on numerous challenging scenarios of reflection and fence removal demonstrate the effectiveness of the proposed method.

CVPR’20: Single-Image HDR Reconstruction by Learning to Reverse the Camera Pipeline

Online live presentation and Q&A times for this paper:

6/16 12:00-14:00 (PDT, Seattle Time) / 6/17 0:00-2:00 (PDT, Seattle Time)

Overview:

Recovering a high dynamic range (HDR) image from a single low dynamic range (LDR) input image is challenging due to missing details in under-/over-exposed regions caused by quantization and saturation of camera sensors. In contrast to existing learning-based methods, our core idea is to incorporate the domain knowledge of the LDR image formation pipeline into our model. We model the HDR-to-LDR image formation pipeline as the (1) dynamic range clipping, (2) non-linear mapping from a camera response function, and (3) quantization. We then propose to learn three specialized CNNs to reverse these steps. By decomposing the problem into specific sub-tasks, we impose effective physical constraints to facilitate the training of individual sub-networks. Finally, we jointly fine-tune the entire model end-to-end to reduce error accumulation. With extensive quantitative and qualitative experiments on diverse image datasets, we demonstrate that the proposed method performs favorably against state-of-the-art single-image HDR reconstruction algorithms.

NTIRE’20: Deploying Image Deblurring across Mobile Devices: A Perspective of Quality and Latency

Online live presentation and Q&A times for this paper:

6/15 17:00-19:00 (PDT, Seattle Time)

Overview:

MediaTek has proposed a practical guideline for better quality-latency trade-off in image deblurring applications on neural network architectures in mobile devices. The paper presents a systematic analysis of the effectiveness of popular network optimization techniques for mobile deployment. It includes observations on quantization and pruning in image deblurring, while also providing an in-depth analysis for both latency and quality. Mediatek’s NeuroPilot SDK and APU (AI Processing Unit) solutions are used for all the experiments and analysis. Using the insights expressed in this paper, MediaTek won the NTIRE 2020 Image Deblurring Challenge event.