lucid-dl/mask-rcnn-resnet-50-fpn

Mask R-CNN (ResNet-50-FPN)

He et al., 2017 — Mask R-CNN (arXiv:1703.06870)

Lucid port of torchvision/MaskRCNN_ResNet50_FPN_Weights.COCO_V1, converted to Lucid-native safetensors.

Available weights

Tag	box mAP	mask mAP	Params	GFLOPs	Size	Source
COCO_V1 (default)	37.9	34.6	44.4M	134.38	169.81 MB	torchvision

Usage

import lucid.models as models
from lucid.models.weights import MaskRCNNResNet50FPNWeights

# default tag
model = models.mask_rcnn_resnet50_fpn(pretrained=True)

# explicit tag (enum or string)
model = models.mask_rcnn_resnet50_fpn(weights=MaskRCNNResNet50FPNWeights.COCO_V1)
model = models.mask_rcnn_resnet50_fpn(pretrained="COCO_V1")

# preprocessing travels with the weights
weights = MaskRCNNResNet50FPNWeights.COCO_V1
preprocess = weights.transforms()
out = model(preprocess(image)[None])
# InstanceSegmentationOutput: class logits + boxes + per-instance masks
logits, boxes, masks = out.logits, out.pred_boxes, out.pred_masks

Conversion

Converted from torchvision/MaskRCNN_ResNet50_FPN_Weights.COCO_V1 via python -m tools.convert_weights mask_rcnn_resnet50_fpn --tag COCO_V1. Key mapping + numerical parity verified against the source.

License

bsd-3-clause — inherited from the original weights.

Citation

@inproceedings{he2017mask,
  title={Mask R-CNN},
  author={He, Kaiming and Gkioxari, Georgia and Doll{\'a}r, Piotr and Girshick, Ross},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  pages={2961--2969},
  year={2017}
}