今天,计算机视觉三大顶会之一CVPR2020接收结果已经公布,一共有1470篇论文被接收,接收率为22%,相比去年降低3个百分点,竞争越来越激烈。
计算机视觉顶会CVPR2020官方今日发布接收论文列表(编号): http://cvpr2020.thecvf.com/sites/default/files/2020-02/accepted_list.txt
1.GhostNet: More Features from Cheap Operations(超越Mobilenet v3的架构) 论文链接:https://arxiv.org/pdf/1911.11907arxiv.org 模型(在ARM CPU上的表现惊人):https://github.com/iamhankai/ghostnetgithub.com
We beat other SOTA lightweight CNNs such as MobileNetV3 and FBNet.
AdderNet: Do We Really Need Multiplications in Deep Learning? (加法神经网络) 在大规模神经网络和数据集上取得了非常好的表现 论文链接:https://arxiv.org/pdf/1912.13200 论文链接:https://github.com/huawei-noah/AdderNet
Frequency Domain Compact 3D Convolutional Neural Networks (3dCNN压缩) 论文链接:https://arxiv.org/pdf/1909.04977 开源代码:https://github.com/huawei-noah/CARSgithub.com
A Semi-Supervised Assessor of Neural Architectures (神经网络精度预测器 NAS)
Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection(NAS 检测) backbone-neck-head一起搜索, 三位一体
CARS: Contunuous Evolution for Efficient Neural Architecture Search (连续进化的NAS) 高效,具备可微和进化的多重优势 论文链接:https://arxiv.org/abs/1909.04977
On Positive-Unlabeled Classification in GAN (PU+GAN)
Learning multiview 3D point cloud registration(3D点云) 论文链接:arxiv.org/abs/2001.05119
Multi-Modal Domain Adaptation for Fine-Grained Action Recognition(细粒度动作识别) 论文链接:arxiv.org/abs/2001.09691
Action Modifiers:Learning from Adverbs in Instructional Video 论文链接:arxiv.org/abs/1912.06617
PolarMask: Single Shot Instance Segmentation with Polar Representation(实例分割建模) 论文链接:arxiv.org/abs/1909.13226 论文解读:https://zhuanlan.zhihu.com/p/84890413 开源代码:https://github.com/xieenze/PolarMask
Rethinking Performance Estimation in Neural Architecture Search(NAS) 由于block wise neural architecture search中真正消耗时间的是performance estimation部分,本文针对 block wise的NAS找到了最优参数,速度更快,且相关度更高。
Distribution Aware Coordinate Representation for Human Pose Estimation(人体姿态估计)
论文链接:arxiv.org/abs/1910.06278 Github:https://github.com/ilovepose/DarkPose 作者团队主页:https://ilovepose.github.io/coco/