UPSNet:A Unified Panoptic Segmentation Network详解（转载）

论文： https://arxiv.org/pdf/1901.03784.pdf
代码： https://github.com/uber-research/UPSNet

UPSNet网络结构

UPSNet总体结构如下图所示，包含RPN、PymaridMaskTarget生成、Semantic Segmentatin Head、Instance Segmentation Head和Panoptic Segmentation Head。

• Backbone Network:特征提取采用ResNet-FPN
• RPN:判断Proposals是否为前景，并生成bbox的坐标变换系数
• PymaridMaskTarget生成：本部分不含计算参数，用于生成前景rois
• Semantic Segmentation Head:包含FCN-Head,语义分割分支
• Instance Segmenation Head:包含RCNN,Mask Branch,预测实例的mask
• Panoptic Segmentation Head:包含SegTerm和MaskTerm，处理整合语义预测和实例的预测
  
  以下逐一介绍各个部件

1. ResNet-FPN

特征的提取采用ResNet后接FPN，FPN的使用充分利用了不同level特征。简图和详细的结构图如下。ResNet每个block的输出（记为$[MathProcessingError]P2P_2$P5​通过一个步长为2，size为1*1的最大池化得到。

整个模块的输出为：

分支名称	channel数	H	W
$[MathProcessingError]FP{N}_{P}2FPN\_P_2$3∗H/i2∗W/i2个anchors 根据RPN得到anchor是前景的得分，对生成的anchor进行排序，并取前N1个 根据RPN得到anchor的box坐标调整系数，调整anchor的box坐标 执行NMS算法，然后取score前N2个，得到该分支的proposals 收集所有分支的proposals，取score前N3的proposals，得到ROIS 该层的输出为ndarray， shape=(2000， 5),第二维的第一位是batch_inx,由于batchsize为1，因此第一位都为0. 3.2 PyramidMaskTarget 该层同样不包含训练参数，两个主要函数add_proposals,sample_rios. add_proposals: 该函数实现了将2000个anchor信息添加到单张图片的roidb标注中。 初始roidb为： { 'boxes':ndarray, shape=(8, 4), 'segms':list, len=8, ‘seg_areas’, list, len=8 'gt_classes':ndarray, len=8, [1 1 1 3 3 3 3 6] 'gt_overlaps':稀疏矩阵， bbox和gt_box的重叠比例，由于当前bbox都是gt_box，只有和其对应类别的处的值为1，shape=(num_box,num_class) ‘is_crowd’: ndarray. len=8 'box_to_gt_ind_map': ndarray, shape=(8,) bbox所对应的gt_box编号 } 1 2 3 4 5 6 7 8 9 通过计算RPN产生的2000个proposals和gt box的重叠比例，修改一字段的值。 boxes，直接将2000个proposals添加到list中 seg_areas，添加2000个0值 gt_classes, 添加2000个0值 gt_overlaps, 添加2000个值，若无重叠则为0，若与一个或多个重叠，取最大值 box_to_gt_ind_map: 添加2000个值，无重叠则为-1， 重叠则为最接近的gt box在gt classes中的序号 sample_rois: 计算bbox和gt box的最大重叠比例，根据阈值进行划分，大于阈值记为前景，小于阈值记为背景。 对前后景rois进行采样，使得总和为rois-per-image, 网络中设置为512个. 计算512个bbox相对于其gt-box的坐标变换系数，即bbox-targets 更改bbox-targets的表示形式，从[class, dx, dy, dh, dw]变成[0,0,0,0,…,dx,dy,dw,dh,…,0,0,0,0],len=num-class * 4, 每四个数对应一个类别，类似于one-hot编码，只在其对应类别处有非零值 为每一个fg-roi制作28*28的mask { 'labels_int32':ndarray, (512,), 若是gt-box，则为对应类别；若非gt-box，则为0， 'rois': (512, 5), 第二维第一位为batch-inx,都为0， 'bbox-targets'：shape（512， 36）， 'bbox-inside-weights': (512, 36),对应bbox-targets,在对应类别处值为1,1,1,1 'bbox-outside-weights': (512, 36), 对应bbox-inside-weights， 在非零值处值为1 'nongt_inds': 不是gtbox的编号， （504，） ->有8个gtbox 'mask_rois': 前景rois, shape(前景rois个数,batchinx+四个系数)=(9，5), 'rois_has_mask_int32': 有mask的rois，即属于前景的rois，shape=(512, )，是则为1，不是为0 'mask_int32': 根据fg-roi宽高比例进行调整的28*28的mask,采用和bbox-targets类似的表示方式，其shape=(前景个数， 类别数×28*28)，在对应属于的类别上，连续的28*28有值（0或1），在不属于的类别上，为-1 } 1 2 3 4 5 6 7 8 9 10 11 4. Sementic Segmentation Head 4.1 FCN-Head FCN-Head即用于生成语义分割的分支。承接FPN的四个分支，每个分支分别经过可变卷积，而后同一scale到H/4、W/4,concat之后再卷积得到每个像素点的在19个类别上的概率。同时计算gt-roi mask内每个像素点的在19个类别上的概率。源码及结构如下： def forward(self, fpn_p2, fpn_p3, fpn_p4, fpn_p5, roi=None): fpn_p2 = self.fcn_subnet(fpn_p2) fpn_p3 = self.fcn_subnet(fpn_p3) fpn_p4 = self.fcn_subnet(fpn_p4) fpn_p5 = self.fcn_subnet(fpn_p5) fpn_p3 <span class="token operator">=</span> F<span class="token punctuation">.</span>interpolate<span class="token punctuation">(</span>fpn_p3<span class="token punctuation">,</span> <span class="token boolean">None</span><span class="token punctuation">,</span> <span class="token number">2</span><span class="token punctuation">,</span> mode<span class="token operator">=</span><span class="token string">'bilinear'</span><span class="token punctuation">,</span> align_corners<span class="token operator">=</span><span class="token boolean">False</span><span class="token punctuation">)</span> fpn_p4 <span class="token operator">=</span> F<span class="token punctuation">.</span>interpolate<span class="token punctuation">(</span>fpn_p4<span class="token punctuation">,</span> <span class="token boolean">None</span><span class="token punctuation">,</span> <span class="token number">4</span><span class="token punctuation">,</span> mode<span class="token operator">=</span><span class="token string">'bilinear'</span><span class="token punctuation">,</span> align_corners<span class="token operator">=</span><span class="token boolean">False</span><span class="token punctuation">)</span> fpn_p5 <span class="token operator">=</span> F<span class="token punctuation">.</span>interpolate<span class="token punctuation">(</span>fpn_p5<span class="token punctuation">,</span> <span class="token boolean">None</span><span class="token punctuation">,</span> <span class="token number">8</span><span class="token punctuation">,</span> mode<span class="token operator">=</span><span class="token string">'bilinear'</span><span class="token punctuation">,</span> align_corners<span class="token operator">=</span><span class="token boolean">False</span><span class="token punctuation">)</span> feat <span class="token operator">=</span> torch<span class="token punctuation">.</span>cat<span class="token punctuation">(</span><span class="token punctuation">[</span>fpn_p2<span class="token punctuation">,</span> fpn_p3<span class="token punctuation">,</span> fpn_p4<span class="token punctuation">,</span> fpn_p5<span class="token punctuation">]</span><span class="token punctuation">,</span> dim<span class="token operator">=</span><span class="token number">1</span><span class="token punctuation">)</span> score <span class="token operator">=</span> self<span class="token punctuation">.</span>score<span class="token punctuation">(</span>feat<span class="token punctuation">)</span> ret <span class="token operator">=</span> <span class="token punctuation">{</span><span class="token string">'fcn_score'</span><span class="token punctuation">:</span> score<span class="token punctuation">,</span> <span class="token string">'fcn_feat'</span><span class="token punctuation">:</span> feat<span class="token punctuation">}</span> <span class="token keyword">if</span> self<span class="token punctuation">.</span>upsample_rate <span class="token operator">!=</span> <span class="token number">1</span><span class="token punctuation">:</span> output <span class="token operator">=</span> F<span class="token punctuation">.</span>interpolate<span class="token punctuation">(</span>score<span class="token punctuation">,</span> <span class="token boolean">None</span><span class="token punctuation">,</span> self<span class="token punctuation">.</span>upsample_rate<span class="token punctuation">,</span> mode<span class="token operator">=</span><span class="token string">'bilinear'</span><span class="token punctuation">,</span> align_corners<span class="token operator">=</span><span class="token boolean">False</span><span class="token punctuation">)</span> ret<span class="token punctuation">.</span>update<span class="token punctuation">(</span><span class="token punctuation">{</span><span class="token string">'fcn_output'</span><span class="token punctuation">:</span> output<span class="token punctuation">}</span><span class="token punctuation">)</span> <span class="token keyword">if</span> roi <span class="token keyword">is</span> <span class="token operator">not</span> <span class="token boolean">None</span><span class="token punctuation">:</span> roi_feat <span class="token operator">=</span> self<span class="token punctuation">.</span>roipool<span class="token punctuation">(</span>feat<span class="token punctuation">,</span> roi<span class="token punctuation">)</span> roi_score <span class="token operator">=</span> self<span class="token punctuation">.</span>score<span class="token punctuation">(</span>roi_feat<span class="token punctuation">)</span> ret<span class="token punctuation">.</span>update<span class="token punctuation">(</span><span class="token punctuation">{</span><span class="token string">'fcn_roi_score'</span><span class="token punctuation">:</span> roi_score<span class="token punctuation">}</span><span class="token punctuation">)</span> <span class="token keyword">return</span> ret

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21

5.Instance Segmentation Head

5.1. RCNN

RCNN分支用于预测512个rois(前后景)的类别和边框

• FPN的四个分支结合RPN生成的rois,使用多层次的RIAlign得到了$[MathProcessingError](512,256,7,7)(512,256,7,7)$ti​表示第i个thing类
  

6.2 MaskTerm

本模块的作用是取出mask分支中每个检测到实例的logit

• mask大小为28*28，将其reshape成对应b-box的大小，将box内的值放到初始化为0的mask　logit对应box位置处

如下图所示.

6.3 计算　panoptic logits

panoptic logits 由三项组成，语义分割部分seg-logits、实例分割部分inst_logits、和未知类预测部分void-logits。其中未知类是由gt-box中随机选出30%作为未知的类别，在本文所用的例子中，原本gt-box共有8个，现在有三个b-box被归为未知类.
void-logits
void-logits根据语义分割实例类部分logits(8,H/4,W/4)最大值减去5个b-box的logits(5,H/4,W/4)的最大值,得到(1, H/4, W/4).void-logits的计算代码如下：

void_logits = torch.max(fcn_output['fcn_score'][:, (config.dataset.num_seg_classes - config.dataset.num_classes + 1):, ...], dim=1, keepdim=True)[0] - torch.max(seg_inst_logits, dim=1, keepdim=True)[0]

 

 
1

inst-logits
inst-logtis则是将MaskTerm和SegTerm得到的mask-logits和seg-ins-logits直接相加。综合了两部分的预测，相当于做了一个模型融合。

inst_logits = seg_inst_logits + mask_logits

 

 
1

seg-logits
seg-logits则是语义分割部分stuffs部分的logtis

panoptic-logits
panoptic-logots直接将三个部分concat在一起。shape=(17, H/4, W/4), 17=11+5+1

panoptic_logits = torch.cat([seg_logits, inst_logits, void_logits], dim=1)

 

 
1

      </div>
      <link href="https://csdnimg.cn/release/phoenix/mdeditor/markdown_views-e44c3c0e64.css" rel="stylesheet">
              </div>