整理:AI算法与图像处理
CVPR2022论文和代码整理:https://github.com/DWCTOD/CVPR2022-Papers-with-Code-Demo
ECCV2022论文和代码整理:https://github.com/DWCTOD/ECCV2022-Papers-with-Code-Demo
标题:FlowFormer: A Transformer Architecture for Optical Flow 论文:https://arxiv.org/abs/2203.16194
主页:https://drinkingcoder.github.io/publication/flowformer/
摘要:本文介绍了光流transformer(FlowFormer),一种基于transformer的神经网络架构,用于学习光流。FlowFormer 对从图像对构建的 4D 成本量进行标记,将cost token 编码到具有新颖潜在空间中的交替组transformer(AGT) 层的成本存储器中,并通过具有动态位置成本查询的循环变换器解码器对成本存储器进行解码 . 在 Sintel 基准测试中,FlowFormer 实现了 1.178 的平均端点误差 (AEPE),与公布的最佳结果 (1.388) 相比,误差减少了 15.1%。此外,FlowFormer 还实现了强大的泛化性能。在没有接受 Sintel 训练的情况下,FlowFormer 在 Sintel 训练集干净通行证上达到 1.00 AEPE,比公布的最佳结果 (1.29) 高出 22.4%。
Embedding contrastive unsupervised features to cluster in- and out-of-distribution noise in corrupted image datasets
Open-world Semantic Segmentation for LIDAR Point Clouds
GraphVid: It Only Takes a Few Nodes to Understand a Video
Target-absent Human Attention
Lottery Ticket Hypothesis for Spiking Neural Networks
Task Discrepancy Maximization for Fine-grained Few-Shot Classification
Egocentric Video-Language Pretraining @ EPIC-KITCHENS-100 Multi-Instance Retrieval Challenge 2022
Aug-NeRF: Training Stronger Neural Radiance Fields with Triple-Level Physically-Grounded Augmentations
PhotoScene: Photorealistic Material and Lighting Transfer for Indoor Scenes
Golfer: Trajectory Prediction with Masked Goal Conditioning MnM Network
DRESS: Dynamic REal-time Sparse Subnets