Increase amount of anchors in FasterRCNN

vision
1

One can alter the sizes of anchors in FasterRCNN by

anchor_sizes = (
      (
          (32,),
          (64,),
          (128,),
          (512,),
          (728,),
      ), # resnet fpn only accepts 5 inputs
)
aspect_ratios = ((0.5, 0.75, 1.0, 1.33, 2.0),) * len(anchor_sizes)
anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)

backbone = resnet_fpn_backbone(
    backbone,
    pretrained=True,
)

model = FasterRCNN(
        backbone,
        num_classes=num_classes,
        rpn_anchor_generator=anchor_generator,
)

In my data with large fluctations of target sizes, this however still leads to inaccurate bounding boxes in large observations.

With mobilenet I can alter the amount of boxes trivially

anchor_sizes = (
    (
        32,
        64,
        128,
        512,
        728,
        882, # 6 box sizes not 5
    ),
)
aspect_ratios = ((0.5, 0.75, 1.0, 1.33, 2.0),) * len(anchor_sizes)
anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)

backbone = mobilenet_backbone(
    "mobilenet_v3_large",
    pretrained=True,
    fpn=False,
)
model = FasterRCNN(
    backbone,
    num_classes=num_classes,
    rpn_anchor_generator=anchor_generator,
)

However, is it possible to use alter ResNet backbone to allow arbitary number of boxes?

For example, this does not work, but can backbone be altered to accept the anchor generator?

anchor_sizes = (
      (
          (32,),
          (64,),
          (128,),
          (512,),
          (728,),
          (882,),
          (1028,),
      ), # here we have seven inputs -> model does not work
)
aspect_ratios = ((0.5, 0.75, 1.0, 1.33, 2.0),) * len(anchor_sizes)
anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)

backbone = resnet_fpn_backbone(
    backbone,
    pretrained=True,
)

model = FasterRCNN(
        backbone,
        num_classes=num_classes,
        rpn_anchor_generator=anchor_generator,
)