pvnet 源github：https://github.com/zju3dv/clean-pvnet
最近在做6D pose estimation时用了pvnet的算法，精度非常不错。想着看看能不能在pvnet上加上目标检测功能，同时判断问题是否存在，原本想着用pvnet的mask做个阈值来判断是否存在物体，结果发现阈值99%了还是无法区分出误检的区域（可能是我背景多样性做的不够）。
整体效果还可以，主要思路还是在resnet18 网络中提最后3层进行目标检测，损失采用yolov4思路。
self.layer5_bbox = nn.Sequential(nn.Conv2d(512, 512, 3, 2, padding=1),nn.BatchNorm2d(512),nn.LeakyReLU(0.1, True))self.layer5_bbox_fc = nn.Conv2d(1024, 3 * (4 + 1 + 1), 3, 1)self.layer6_bbox = nn.Sequential(nn.Conv2d(512, 512, 3, 2, padding=2),nn.BatchNorm2d(512),nn.LeakyReLU(0.1, True))self.layer6_bbox_up = nn.UpsamplingBilinear2d(size=(38, 38))self.layer6_bbox_fc = nn.Conv2d(1536, 3 * (4 + 1 + 1), 3, 1)self.layer7_bbox = nn.Sequential(nn.Conv2d(512, 1024, 3, 2, padding=2),nn.BatchNorm2d(1024),nn.LeakyReLU(0.1, True))self.layer7_bbox_up = nn.UpsamplingBilinear2d(size=(20, 20))self.layer7_bbox_fc = nn.Conv2d(1024, 3 * (4 + 1 + 1), 3, 1)def forward(self, x, feature_alignment=False):x2s, x4s, x8s, x16s, x32s, xfc = self.resnet18_8s(x)bbox_1 = self.layer5_bbox(x32s)bbox_2 = self.layer6_bbox(bbox_1)bbox_2_up = self.layer6_bbox_up(bbox_2)bbox_3 = self.layer7_bbox(bbox_2)bbox_3_up = self.layer7_bbox_up(bbox_3)bbox_1_fc = self.layer5_bbox_fc(torch.cat([bbox_1, bbox_2_up], 1))bbox_2_fc = self.layer6_bbox_fc(torch.cat([bbox_2, bbox_3_up], 1))bbox_3_fc = self.layer7_bbox_fc(bbox_3)# x2s, x4s, x8s, _, _, xfcfm = self.conv8s(torch.cat([xfc, x8s], 1)) # 结合ｒｅｓｎｅｔ１８的 x8s（第三层） xfc（最后一层）然后用于上采样fm = self.up8sto4s(fm) # 第一层 上采样一倍# if fm.shape[2] == 136:# fm = nn.functional.interpolate(fm, (135, 180), mode='bilinear', align_corners=False)fm = self.conv4s(torch.cat([fm, x4s], 1))fm = self.up4sto2s(fm) # 第二层 上采样一倍fm = self.conv2s(torch.cat([fm, x2s], 1))fm = self.up2storaw(fm) # 第三层 上采样一倍x = self.convraw(torch.cat([fm, x], 1))seg_pred = x[:, :2, :, :] # 前２个维度是分割# if self.seg_dim > 2:# bounding_box = x[:, 2:self.seg_dim, :, :] # 前２个维度是分割# else:# bounding_box = Nonever_pred = x[:, self.seg_dim:, :, :] # 后１８个维度是 ８个关键点＋１个中心点, vector-field# ret = {'seg': seg_pred, 'vertex': ver_pred} # @!!!@ret = {'seg': seg_pred,'bounding_box': [bbox_1_fc, bbox_2_fc, bbox_3_fc],'vertex': ver_pred}if not self.training:with torch.no_grad():self.decode_keypoint(ret)return ret