deeplabv3+ 笔记

Deeplab v3+
Abstract
(1)Deeplab v3+ adds a decoder modeule to refine the segmentation result along object boundaries.
(2)Explore the Xception model and apply the depthwise separable convolution to both Atrous Spatial Pyramid Pooling and decoder modules
(3)好句：encoder multi-scale contextual information by probing the incoming features with filters or pooling operations at multiple rates and multiple effective fields-of-view. Network can capture sharper object boundaries by recovering the spatial information
Introduce
(1)in order to capture the contextual information at multiple scales, deeplabv3 applies several parallel atrous conv with different rates(ASPP)
(2)Information related to object boundaries miss due to pool or conv with striding operations, So apply atrous conv can alleviate
(3)Deeplabv3+ contains rich segmentation information from encoder, (ASPP) , while recovery the object boundaries from decoder既能通过encoder的ASPP结构得到丰富的上下文信息，又能通过decoder来恢复物体的边缘空间信息。

Related work
(1)Spatial pyramind pooling: capture rich contextual information
(2)Encoder-decoder: recover the spatial information,refine the segmentation result
(3)Depthwise seperation conv (or group conv): reduce the computation cost and number of params
Method

(1)Image pooling是GAP， output是16倍的，Low-Level-Feature是resnet的conv2的特征图。
(2)Xception调整 backbone 中max pool instead of conv with striding operations， bn+relu after each 3x3 depthwise conv（like mobilenet）

注:图片全部来源于Deeplabv3+论文原文。