Rethinking Atrous Convolution fo

Introduction

语义分割的两个挑战
The first one is the reducedfeature resolution caused by consecutive pooling operationsor convolution striding
解决方法：
带孔卷积

Another difficulty comes from the existence of object sat multiple scales
解决方法：
图像金字塔
Encoder-Decoder
DenseCRF
空间金字塔池化

Related Work

Image pyramid
Encoder-decoder
Context module
Spatial pyramid pooling
Atrous convolution

Methods

Atrous Convolution for Dense Feature Extraction
用带孔卷积提取特征

空洞卷积详解
Multi-Scale Context Aggregation by Dilated Convolution

Going Deeper with Atrous Convolution

级联结构

Multi-grid Method
We adopt different atrous rates within block4 toblock7 in the proposed model. In particular, we define as MultiGrid= (r1, r2, r3) the unit rates for the three convo-lutional layers within block4 to block7

ASPP

In the extreme case where the rate value is close to the feature map size, the3×3 filter, instead of capturing the whole image context, degenerates to a simple 1×1 filter since only the center filter weight is effective.

Valid Weight:he weights that are applied to the valid fea-ture region, instead of padded zeros

We apply global average pooling on the last feature map of the model, feed the resulting image-level features to a 1×1 convolution with 256 filters (and batch normalization [38]), and then bilinearly upsample the feature to the desired spatial dimension.

全局平均池化

image_level_features = tf.reduce_mean(net, [1, 2], name='image_level_global_pool', keepdims=True)
image_level_features = slim.conv2d(image_level_features, depth, [1, 1], 
    scope="image_level_conv_1x1",activation_fn=None)
image_level_features = tf.image.resize_bilinear(image_level_features, (feature_map_size[1], feature_map_size[2]))

并联结构

Experinment