RAFT
RAFT (Recurrent All-Pairs Field Transforms) is a deep learning-based optical flow estimation model. It constructs all-pairs correlation volumes between feature maps and iteratively updates flow fields using a recurrent unit, achieving high accuracy on standard optical flow benchmarks.
Core Facts
Timeline (last 90 days)
Experimental results show that RAFT reduces training input by 35.23% on average, and by up to 97.11%.
The paper "RAFT: Redundancy-Aware Input Filtering for Temporal Graph Neural Network Training" proposes RAFT, a redundancy-aware input filtering framework for TGNN training.
RAFT contains two filters: the root-event filter removes selected prediction targets before training; the support-event filter excludes selected historical context during sampling.
RAFT uses the magnitude of node memory change to judge the importance of an event: a small memory change means the event has limited impact on subsequent predictions and can serve as a basis for redundancy filtering.
Under multiple evaluation settings, filtering out historical events related to stable memories can maintain similar training loss trajectories.
The researchers evaluated RAFT on three TGNN models and five datasets.
The core process of training a model is to iteratively update parameters in the direction of the loss function's gradient via gradient descent.
Deep learning-based optical flow estimation methods such as RAFT and FlowNet have achieved significant breakthroughs in accuracy and speed, making real-time optical flow frame interpolation possible.
光流估计（通过 RAFT、GMFlow 等算法）用于量化视频运动信息，光流均值接近零表示静态场景，方差过大意味着剧烈场景切换或抖动，这类数据不适合训练
光流计算连续两帧之间每个像素的位移向量形成稠密运动场，经典算法包括Lucas-Kanade和Horn-Schunck，RAFT等深度学习方法提升了精度
2 more timeline events
All Facts (12)
The core process of training a model is to iteratively update parameters in the direction of the loss function's gradient via gradient descent.
75%UnverifiedUnder multiple evaluation settings, filtering out historical events related to stable memories can maintain similar training loss trajectories.
50%UnverifiedExperimental results show that RAFT reduces training input by 35.23% on average, and by up to 97.11%.
50%UnverifiedRAFT contains two filters: the root-event filter removes selected prediction targets before training; the support-event filter excludes selected historical context during sampling.
50%UnverifiedRAFT uses the magnitude of node memory change to judge the importance of an event: a small memory change means the event has limited impact on subsequent predictions and can serve as a basis for redundancy filtering.
50%UnverifiedThe researchers evaluated RAFT on three TGNN models and five datasets.
50%UnverifiedDeep learning-based optical flow estimation methods such as RAFT and FlowNet have achieved significant breakthroughs in accuracy and speed, making real-time optical flow frame interpolation possible.
50%Unverified光流估计（通过 RAFT、GMFlow 等算法）用于量化视频运动信息，光流均值接近零表示静态场景，方差过大意味着剧烈场景切换或抖动，这类数据不适合训练
50%Unverified光流计算连续两帧之间每个像素的位移向量形成稠密运动场，经典算法包括Lucas-Kanade和Horn-Schunck，RAFT等深度学习方法提升了精度
50%Unverified经典光流算法Lucas-Kanade基于局部约束，Horn-Schunck基于全局约束，都依赖亮度恒定假设；FlowNet（2015）首次用CNN端到端估计光流，RAFT（2020）通过全对相关体积改进
50%UnverifiedFlowNet于2015年首次用卷积神经网络端到端估计光流，RAFT于2020年通过全对相关体积并迭代优化大幅提升光流估计精度
50%UnverifiedThe paper "RAFT: Redundancy-Aware Input Filtering for Temporal Graph Neural Network Training" proposes RAFT, a redundancy-aware input filtering framework for TGNN training.
50%