LibreFCNr101

Torchvision FCN with a dilated ResNet-101 backbone, repackaged for LibreYOLO. The 2015 FCN work established end-to-end pixels-to-pixels prediction, but this checkpoint uses torchvision's later ResNet graph. It is not the original paper's VGG-based FCN-8s skip-fusion architecture.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreFCNr101.pt")
result = model.predict("image.jpg")
mask = result.semantic_mask.data

Source

Derived from pytorch/vision at commit 336d36e8db990a905498c73933e35231876e28bc. Copyright (c) Soumith Chintala 2016 and the torchvision contributors. The source implementation is BSD-3-Clause.

Official checkpoint: fcn_resnet101_coco-7ecb50ca.pth Official checkpoint bytes: 217,800,805 SHA-256: 7ecb50ca17844860a70d5ed0c748d997cf8adb62932abaa0233430c68594d749

Torchvision reports COCO-val2017-VOC-labels mIoU 63.7 and pixel accuracy 91.9.

Categories

The 21 output channels are __background__, aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, diningtable, dog, horse, motorbike, person, pottedplant, sheep, sofa, train, tvmonitor.

Modifications

Checkpoint metadata was added for LibreYOLO's v1.0 schema. Learned tensors and state-dict keys are unchanged, including the primary and auxiliary heads. The native LibreYOLO graph strict-loads the official state dict and both dense-logit outputs are bit-exact against torchvision. See weights/convert_fcn_weights.py in the LibreYOLO source repository.

License

The checkpoint publisher did not attach a separate per-object license file. This mirror applies the releasing project's BSD-3-Clause license on an implied, not publisher-confirmed, basis. Torchvision warns that pretrained models may have their own licenses or terms derived from training data and that users must determine whether they have permission for their use case. See LICENSE and NOTICE.