Download presentation
Presentation is loading. Please wait.
Published byValerio Pippi Modified 5年之前
3
HRNet 保持高分辨率 不同分辨率之间进行信息交换(exchange) Exchange Unit HRNet
Exchange Block
4
Experiment results
5
Ablation study
6
HRNet v2 将HRNet用在分割和检测任务上
7
Modification:多个分辨率的输出concat
8
Experiment Semantic segmentation Object detection
Face landmark detection
10
对不同位置的attention-map进行可视化
12
我做的可视化
13
Motivation Non-local所有位置的attention-map几乎相同,可以简化
简化后的non-local跟SE Block惊人的结构相似 Non-local和SENet可以结合
14
简化NL Block
15
Simplified NL + SE = Global context block
16
Ablation study on COCO Validation
Non-local论文结果
17
Result on Kinetics Non-local论文结果
19
Motivation Self-attention计算每个位置的attention都需要所有位置参与,计算 量庞大(O(n^2)) 且没必要
本文提出轻量级的lightweight convolution和dynamic convolution
21
Lightweight convolution
有weight share的depthwise convolution;分H组,kernel size为k; 卷积核参数在k上进行softmax Gated linear unit,输入2d,其中d维用来做sigmoid后与另外d维相乘(便于优化梯度)
22
Dynamic Convolution 卷积核的参数并不取决于entire context,仅是当前的time-step的函数
24
Background Attention过程中包括了key和query,起作用的包括:key content, query content以及relative position
25
Spatial attention mechanisms
Generalized attention formulation Transformer attention 第m个attention head的attention weight
26
Deformable convolution
Regular convolution Deformable convolution Dynamic convolution 卷积offset Deformation offset 双线性插值
27
𝜀 1 = z 𝑞 𝑇 𝑈 𝑚 𝑇 𝑉 𝑚 𝐶 𝑥 𝑘 𝐶× 𝑁 𝑠 𝑁 𝑠 ×𝐶 𝐶×𝐶 𝐶×𝐶 𝑂(𝑁 𝑠 𝐶 2 ) 𝑂( 𝑁 𝑠 𝐶 2 ) 𝑁 𝑠 ×𝐶 𝐶× 𝑁 𝑠 𝑂( 𝑁 𝑠 2 𝐶) 𝑁 𝑠 × 𝑁 𝑠 总复杂度:𝑂( 𝑁 𝑠 2 𝐶+ 𝑁 𝑠 𝐶 2 )
28
Experiment Object detection/semantic segmentation NMT
29
Disentangle Transformer attention module
Similar presentations