mmdetection中的backbone为了添加FPN进行多尺度的检测，backbone的输入是４个stage的feature map的tensor，详见resnet.py源码中的Example:

    1.         >>> from mmdet.models import ResNet
    2.         >>> import torch
    3.         >>> self = ResNet(depth=18)
    4.         >>> self.eval()
    5.         >>> inputs = torch.rand(1, 3, 32, 32)
    6.         >>> level_outputs = self.forward(inputs)
    7.         >>> for level_out in level_outputs:
    8.         ...     print(tuple(level_out.shape))
    9.         (1, 64, 8, 8)
    10.         (1, 128, 4, 4)
    11.         (1, 256, 2, 2)
    12.         (1, 512, 1, 1)

    Backbone(以resnet.py为例)中的forward函数return的是一个tuple，并且包含四个不同stage的feature map:

    1. class ResNet(nn.Module):
    3. ........
    4. ........
    5. ........
    6. ........
    8.     def forward(self, x):
    9.         x = self.conv1(x)
    10.         x = self.norm1(x)
    11.         x = self.relu(x)
    12.         x = self.maxpool(x)
    13.         outs = []
    14.         for i, layer_name in enumerate(self.res_layers):
    15.             res_layer = getattr(self, layer_name)
    16.             x = res_layer(x)
    17.             if i in self.out_indices:
    18.                 outs.append(x)
    19.         return tuple(outs)

    所以在测试的过程中就很好的就行特征图可视化的操作。

    首先在mmdet/models/detectors/two_stage.py中找到test的函数：

    1. class TwoStageDetector(BaseDetector, RPNTestMixin, BBoxTestMixin,
    2.                        MaskTestMixin):
    3. ........
    4. ........
    5. ........
    6. ........
    8.     def extract_feat(self, img):
    9.         """Directly extract features from the backbone+neck
    10.         """
    11.         x = self.backbone(img)
    12.         # 此时在这里输出的就是backbone的４个stage的feature map
    13.         if self.with_neck: # 如果有neck,比如FPN，最终的输出就是4个stage经过FPN处理后的feature map
    14.             x = self.neck(x)
    15.         return x
    18. ........
    19. ........
    20. ........
    21. ........
    23.     def simple_test(self, img, img_meta, proposals=None, rescale=False):
    24.         """Test without augmentation."""
    25.         assert self.with_bbox, "Bbox head must be implemented."
    27.         x = self.extract_feat(img)
    28.         # print(x[0].shape, len(x))
    31.         if proposals is None:
    32.             proposal_list = self.simple_test_rpn(x, img_meta,
    33.                                                  self.test_cfg.rpn)
    34.         else:
    35.             proposal_list = proposals
    37.         det_bboxes, det_labels = self.simple_test_bboxes(
    38.             x, img_meta, proposal_list, self.test_cfg.rcnn, rescale=rescale)
    39.         bbox_results = bbox2result(det_bboxes, det_labels,
    40.                                    self.bbox_head.num_classes)
    42.         if not self.with_mask:
    43.             return bbox_results    #   modified origin : return bbox_results
    44.             # return bbox_results
    45.         else:
    46.             segm_results = self.simple_test_mask(
    47.                 x, img_meta, det_bboxes, det_labels, rescale=rescale)
    48.             return bbox_results, segm_results

    所以如果要将feature map输出，就要更改这两个函数的返回值，首先：

    1.     def extract_feat(self, img):
    2.         """Directly extract features from the backbone+neck
    3.         """
    4.         x = self.backbone(img)
    5.         if self.with_neck:
    6.             x = self.neck(x)
    7.         return x
    9.      # change the code to 
    10.     def extract_feat(self, img):
    11.         """Directly extract features from the backbone+neck
    12.         """
    13.         x = self.backbone(img)
    14.         if self.with_neck:
    15.             _x = self.neck(x)
    16.         return x, _x
    18.     # 也可以在x = self.backbone(img)下面一行将这个元组中的tensor转变为numpy形式进行保存，
    19.     # 在通过opencv, PIL进行图像可视化

    不让原来backbone中输出的feature map的变量被覆盖掉，所以可以使用不同的变量名进行赋值操作。

    原来的simple_test函数是只返回图片中检测到物体的坐标，这里我们稍微做一下修改：

    1.     def simple_test(self, img, img_meta, proposals=None, rescale=False):
    2.         """Test without augmentation."""
    3.         assert self.with_bbox, "Bbox head must be implemented."
    5.         # origin code : x = self.extract_feat(img)
    6.         x, _x = self.extract_feat(img)
    7.         '''
    8.         x : backbone中四个stage的feature map的tensor
    9.         _x : 经过ＦＰＮ后4个stage的feature map的tensor
    10.         '''
    11.         # print(x[0].shape, len(x))
    14.         if proposals is None:
    15.             proposal_list = self.simple_test_rpn(x, img_meta,
    16.                                                  self.test_cfg.rpn)
    17.         else:
    18.             proposal_list = proposals
    20.         det_bboxes, det_labels = self.simple_test_bboxes(
    21.             x, img_meta, proposal_list, self.test_cfg.rcnn, rescale=rescale)
    22.         bbox_results = bbox2result(det_bboxes, det_labels,
    23.                                    self.bbox_head.num_classes)
    25.         if not self.with_mask:
    26.             '''
    27.             在这里加上x, _x的返回值就行！！！！
    28.             '''
    29.             return bbox_results, x      #   modified origin : return bbox_results
    30.             # return bbox_results
    31.         else:
    32.             segm_results = self.simple_test_mask(
    33.                 x, img_meta, det_bboxes, det_labels, rescale=rescale)
    34.             return bbox_results, segm_results