目标检测之RPN网络（Feature Pyramid Networks for Object Detection）

最近读了几篇WSDN的文章，有一篇的paper用到了RPN网络的一些思想，因此去拜读一下RPN这篇paper。这篇paper的attribution就是取长补短的思想，在卷积神经网络中，网络层数越浅，语意特征不明显，但是图像的特征却比较明显；网络层数越深，语义特征却越明显；这个其实从卷积神经网络的可视化都是可以看出来的。因此这篇paper想用的思想就是将深层的抽象特征与浅层丰富的几何信息结合到一起，这样生成的feature含有的信息就比较丰沛。尤其对于目标检测来说，深层的网络虽然具有很高的语义信息，但是Feature map比较小，含有的几何信息不是很多，不利用小目标的检测；浅层网络恰恰相反，没有比较高的语义信息，但是却含有比较多的几何信息，分辨率比较高。因此如果将这两者结合起来做目标检测的工作将会大大的提升小目标的检测。

1、Background

（a）十分耗时，计算速度比较慢。

（b）速度提升了，但是只有最后一层的高层语义特征，让低层的信息可能会丢失。

（c）虽然SSD中使用了多尺度的feature map但是没有使用低层的feature map

（d）既是多尺度，而且还将高层的语义特征和浅层的语义特征结合起来。通过自底向上，自顶向下，横向连接的方式。

将高层语义特征进行2倍上采样得到的特征和浅层语义的特征结合起来。其中1*1的卷积是为了减小通道数。

Bottom-up pathway：将输入的图像不断的进行卷积，最终得到最终的特征图。

Top-down pathway and lateral connections：将最后的特征图与下一层的特征图结合起来，得到更丰富的语义特征称之为自顶向下横向连接。

2、FPN网络

Attribution
- 端到端的网络，不管在测试还是训练
- 更高的准确率
- 提出RPN网络
- 没有增加测试时间提升了网络的性能
Result
- 将faster rcnn的骨干网络换成RPN，mAP从51.7%->56.9%

3、代码

RPN核心代码

import random
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable, gradcheck
from torch.autograd.gradcheck import gradgradcheck
import torchvision.models as models
from torch.autograd import Variable
import numpy as np
import torchvision.utils as vutils
from model.utils.config import cfg
from model.rpn.rpn_fpn import _RPN_FPN
from model.roi_pooling.modules.roi_pool import _RoIPooling
from model.roi_crop.modules.roi_crop import _RoICrop
from model.roi_align.modules.roi_align import RoIAlignAvg
from model.rpn.proposal_target_layer import _ProposalTargetLayer
from model.utils.net_utils import _smooth_l1_loss, _crop_pool_layer, _affine_grid_gen, _affine_theta
import time
import pdbclass _FPN(nn.Module):""" FPN """def __init__(self, classes, class_agnostic):super(_FPN, self).__init__()self.classes = classesself.n_classes = len(classes)self.class_agnostic = class_agnostic# lossself.RCNN_loss_cls = 0self.RCNN_loss_bbox = 0self.maxpool2d = nn.MaxPool2d(1, stride=2)# define rpnself.RCNN_rpn = _RPN_FPN(self.dout_base_model)self.RCNN_proposal_target = _ProposalTargetLayer(self.n_classes)# NOTE: the original paper used pool_size = 7 for cls branch, and 14 for mask branch, to save the# computation time, we first use 14 as the pool_size, and then do stride=2 pooling for cls branch.self.RCNN_roi_pool = _RoIPooling(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1.0/16.0)self.RCNN_roi_align = RoIAlignAvg(cfg.POOLING_SIZE, cfg.POOLING_SIZE, 1.0/16.0)self.grid_size = cfg.POOLING_SIZE * 2 if cfg.CROP_RESIZE_WITH_MAX_POOL else cfg.POOLING_SIZEself.RCNN_roi_crop = _RoICrop()def _init_weights(self):def normal_init(m, mean, stddev, truncated=False):"""weight initalizer: truncated normal and random normal."""# x is a parameterif truncated:m.weight.data.normal_().fmod_(2).mul_(stddev).add_(mean) # not a perfect approximationelse:m.weight.data.normal_(mean, stddev)m.bias.data.zero_()# custom weights initialization called on netG and netDdef weights_init(m, mean, stddev, truncated=False):classname = m.__class__.__name__if classname.find('Conv') != -1:m.weight.data.normal_(0.0, 0.02)m.bias.data.fill_(0)elif classname.find('BatchNorm') != -1:m.weight.data.normal_(1.0, 0.02)m.bias.data.fill_(0)normal_init(self.RCNN_toplayer, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_smooth1, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_smooth2, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_smooth3, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_latlayer1, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_latlayer2, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_latlayer3, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_rpn.RPN_Conv, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_rpn.RPN_cls_score, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_rpn.RPN_bbox_pred, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_cls_score, 0, 0.01, cfg.TRAIN.TRUNCATED)normal_init(self.RCNN_bbox_pred, 0, 0.001, cfg.TRAIN.TRUNCATED)weights_init(self.RCNN_top, 0, 0.01, cfg.TRAIN.TRUNCATED)def create_architecture(self):self._init_modules()self._init_weights()def _upsample_add(self, x, y):'''Upsample and add two feature maps.Args:x: (Variable) top feature map to be upsampled.y: (Variable) lateral feature map.Returns:(Variable) added feature map.Note in PyTorch, when input size is odd, the upsampled feature mapwith `F.upsample(..., scale_factor=2, mode='nearest')`maybe not equal to the lateral feature map size.e.g.original input size: [N,_,15,15] ->conv2d feature map size: [N,_,8,8] ->upsampled feature map size: [N,_,16,16]So we choose bilinear upsample which supports arbitrary output sizes.'''_,_,H,W = y.size()return F.upsample(x, size=(H,W), mode='bilinear') + ydef _PyramidRoI_Feat(self, feat_maps, rois, im_info):''' roi pool on pyramid feature maps'''# do roi pooling based on predicted roisimg_area = im_info[0][0] * im_info[0][1]h = rois.data[:, 4] - rois.data[:, 2] + 1w = rois.data[:, 3] - rois.data[:, 1] + 1roi_level = torch.log(torch.sqrt(h * w) / 224.0)roi_level = torch.round(roi_level + 4)roi_level[roi_level < 2] = 2roi_level[roi_level > 5] = 5# roi_level.fill_(5)if cfg.POOLING_MODE == 'crop':# pdb.set_trace()# pooled_feat_anchor = _crop_pool_layer(base_feat, rois.view(-1, 5))# NOTE: need to add pyrmaidgrid_xy = _affine_grid_gen(rois, base_feat.size()[2:], self.grid_size)grid_yx = torch.stack([grid_xy.data[:,:,:,1], grid_xy.data[:,:,:,0]], 3).contiguous()roi_pool_feat = self.RCNN_roi_crop(base_feat, Variable(grid_yx).detach())if cfg.CROP_RESIZE_WITH_MAX_POOL:roi_pool_feat = F.max_pool2d(roi_pool_feat, 2, 2)elif cfg.POOLING_MODE == 'align':roi_pool_feats = []box_to_levels = []for i, l in enumerate(range(2, 6)):if (roi_level == l).sum() == 0:continueidx_l = (roi_level == l).nonzero().squeeze()box_to_levels.append(idx_l)scale = feat_maps[i].size(2) / im_info[0][0]feat = self.RCNN_roi_align(feat_maps[i], rois[idx_l], scale)roi_pool_feats.append(feat)roi_pool_feat = torch.cat(roi_pool_feats, 0)box_to_level = torch.cat(box_to_levels, 0)idx_sorted, order = torch.sort(box_to_level)roi_pool_feat = roi_pool_feat[order]elif cfg.POOLING_MODE == 'pool':roi_pool_feats = []box_to_levels = []for i, l in enumerate(range(2, 6)):if (roi_level == l).sum() == 0:continueidx_l = (roi_level == l).nonzero().squeeze()box_to_levels.append(idx_l)scale = feat_maps[i].size(2) / im_info[0][0]feat = self.RCNN_roi_pool(feat_maps[i], rois[idx_l], scale)roi_pool_feats.append(feat)roi_pool_feat = torch.cat(roi_pool_feats, 0)box_to_level = torch.cat(box_to_levels, 0)idx_sorted, order = torch.sort(box_to_level)roi_pool_feat = roi_pool_feat[order]return roi_pool_featdef forward(self, im_data, im_info, gt_boxes, num_boxes):batch_size = im_data.size(0)im_info = im_info.datagt_boxes = gt_boxes.datanum_boxes = num_boxes.data# feed image data to base model to obtain base feature map# Bottom-up#自底向上c1 = self.RCNN_layer0(im_data)c2 = self.RCNN_layer1(c1)c3 = self.RCNN_layer2(c2)c4 = self.RCNN_layer3(c3)c5 = self.RCNN_layer4(c4)# Top-down#自顶向下+横向连接p5 = self.RCNN_toplayer(c5)p4 = self._upsample_add(p5, self.RCNN_latlayer1(c4))p4 = self.RCNN_smooth1(p4)p3 = self._upsample_add(p4, self.RCNN_latlayer2(c3))p3 = self.RCNN_smooth2(p3)p2 = self._upsample_add(p3, self.RCNN_latlayer3(c2))p2 = self.RCNN_smooth3(p2)p6 = self.maxpool2d(p5)rpn_feature_maps = [p2, p3, p4, p5, p6]mrcnn_feature_maps = [p2, p3, p4, p5]#通过经过RPN网络得到proposalsrois, rpn_loss_cls, rpn_loss_bbox = self.RCNN_rpn(rpn_feature_maps, im_info, gt_boxes, num_boxes)# if it is training phrase, then use ground trubut bboxes for refiningif self.training:roi_data = self.RCNN_proposal_target(rois, gt_boxes, num_boxes)rois, rois_label, gt_assign, rois_target, rois_inside_ws, rois_outside_ws = roi_data## NOTE: additionally, normalize proposals to range [0, 1],#        this is necessary so that the following roi pooling#        is correct on different feature maps# rois[:, :, 1::2] /= im_info[0][1]# rois[:, :, 2::2] /= im_info[0][0]rois = rois.view(-1, 5)rois_label = rois_label.view(-1).long()gt_assign = gt_assign.view(-1).long()pos_id = rois_label.nonzero().squeeze()gt_assign_pos = gt_assign[pos_id]rois_label_pos = rois_label[pos_id]rois_label_pos_ids = pos_idrois_pos = Variable(rois[pos_id])rois = Variable(rois)rois_label = Variable(rois_label)rois_target = Variable(rois_target.view(-1, rois_target.size(2)))rois_inside_ws = Variable(rois_inside_ws.view(-1, rois_inside_ws.size(2)))rois_outside_ws = Variable(rois_outside_ws.view(-1, rois_outside_ws.size(2)))else:## NOTE: additionally, normalize proposals to range [0, 1],#        this is necessary so that the following roi pooling#        is correct on different feature maps# rois[:, :, 1::2] /= im_info[0][1]# rois[:, :, 2::2] /= im_info[0][0]rois_label = Nonegt_assign = Nonerois_target = Nonerois_inside_ws = Nonerois_outside_ws = Nonerpn_loss_cls = 0rpn_loss_bbox = 0rois = rois.view(-1, 5)pos_id = torch.arange(0, rois.size(0)).long().type_as(rois).long()rois_label_pos_ids = pos_idrois_pos = Variable(rois[pos_id])rois = Variable(rois)# pooling features based on rois, output 14x14 map#对于不同尺度的feature进行分开roi poolingroi_pool_feat = self._PyramidRoI_Feat(mrcnn_feature_maps, rois, im_info)# feed pooled features to top model#将最后得到的特征向量映射成4096维度pooled_feat = self._head_to_tail(roi_pool_feat)#之后在计算分类和回归# compute bbox offsetbbox_pred = self.RCNN_bbox_pred(pooled_feat)if self.training and not self.class_agnostic:# select the corresponding columns according to roi labelsbbox_pred_view = bbox_pred.view(bbox_pred.size(0), int(bbox_pred.size(1) / 4), 4)bbox_pred_select = torch.gather(bbox_pred_view, 1, rois_label.long().view(rois_label.size(0), 1, 1).expand(rois_label.size(0), 1, 4))bbox_pred = bbox_pred_select.squeeze(1)# compute object classification probabilitycls_score = self.RCNN_cls_score(pooled_feat)cls_prob = F.softmax(cls_score)RCNN_loss_cls = 0RCNN_loss_bbox = 0if self.training:# loss (cross entropy) for object classificationRCNN_loss_cls = F.cross_entropy(cls_score, rois_label)# loss (l1-norm) for bounding box regressionRCNN_loss_bbox = _smooth_l1_loss(bbox_pred, rois_target, rois_inside_ws, rois_outside_ws)rois = rois.view(batch_size, -1, rois.size(1))cls_prob = cls_prob.view(batch_size, -1, cls_prob.size(1))bbox_pred = bbox_pred.view(batch_size, -1, bbox_pred.size(1))if self.training:rois_label = rois_label.view(batch_size, -1)return rois, cls_prob, bbox_pred, rpn_loss_cls, rpn_loss_bbox, RCNN_loss_cls, RCNN_loss_bbox, rois_label

4、总结

这篇paper发表的时间比较早了，现在很多目标检测的任务已经引用到了RPN的思想，就好比把大象放到冰箱里，可以分成两个步骤，但是每个步骤的基础知识原理如果搞清楚了，从最基本的网络原理去解决问题，那么对于模型性能的提升会有很大帮助的。

代码来源：https://github.com/jwyang/fpn.pytorch

参考博客：https://www.cnblogs.com/hansjorn/p/12510888.html

论文：https://arxiv.org/abs/1612.03144