How to increase Mask R-CNN recall

Bernd · October 26, 2021, 9:03pm

Hello PyTorchers
I am currently training a PyTorch Mask R-CNN based on ResNet50 backbone.

pytorch/vision/blob/main/torchvision/models/detection/mask_rcnn.py

from collections import OrderedDict

from torch import nn
from torchvision.ops import MultiScaleRoIAlign

from ..._internally_replaced_utils import load_state_dict_from_url
from ...ops import misc as misc_nn_ops
from ..resnet import resnet50
from ._utils import overwrite_eps
from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
from .faster_rcnn import FasterRCNN

__all__ = [
    "MaskRCNN",
    "maskrcnn_resnet50_fpn",
]


class MaskRCNN(FasterRCNN):
    """

This file has been truncated. show original

During my training session I noticed that the precision is very good, going up to 94%, but the recall only up to 79%. For my use case it is very important that I detect all objects, while the actual classification is less important.
There are tons of parameters in the MaskRCNN base class, and I wonder if any of them could be changed to move toward a higher recall in favor of some precision.

I tried playing around with box_fg_iou_thresh and box_bg_iou_thresh, but without any noticable change in recall or precision.
Has anyone a suggestion which parameters could be used to increase recall?

Regards,
Bernd