I need to know what feature location every anchor box is corresponding to, so I modify the torchvision.models.detection.rpn.AnchorGenerator in this way:
class AnchorGeneratorMP(AnchorGenerator):
def __init__(
self,
sizes=(128, 256, 512),
aspect_ratios=(0.5, 1.0, 2.0),
):
super(AnchorGeneratorMP, self).__init__(sizes, aspect_ratios)
def forward(self, image_list, feature_maps):
grid_sizes = tuple([feature_map.shape[-2:] for feature_map in feature_maps])
image_size = image_list.tensors.shape[-2:]
strides = tuple((image_size[0] / g[0], image_size[1] / g[1]) for g in grid_sizes)
self.set_cell_anchors(feature_maps[0].device)
anchors_over_all_feature_maps = self.cached_grid_anchors(grid_sizes, strides)
anchors = []
for i, (image_height, image_width) in enumerate(image_list.image_sizes):
anchors_in_image = []
for anchors_per_feature_map in anchors_over_all_feature_maps:
anchors_in_image.append(anchors_per_feature_map)
anchors.append(anchors_in_image)
# here is the only place where I modified, except the returns
anchor_len = [len(anchor) for anchor in anchors_over_all_feature_maps]
anchors = [t.cat(anchors_per_image) for anchors_per_image in anchors]
return anchors, anchor_len
Conforming to the FasterRCNN implemented by pytorch, the sizes and aspect_ratios are
anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
So it is supposed to output 15 anchors per feature location. What makes me confused is that for every anchor_len, it is only 3 times the total locations of the corresponding feature. Where am I missing?