LibreFCOS3D

Byte-identical mirror of the official FCOS3D R101-DCN nuScenes checkpoint (finetuned variant), for use with LibreYOLO's LibreFCOS3D inference adapter.

⚠️ NON-COMMERCIAL WEIGHTS, AND AN IMPLIED REDISTRIBUTION BASIS

These weights are not covered by LibreYOLO's MIT license. Two separate restrictions apply, and both bind you, the downloader:

1. Training data. This checkpoint was trained solely on nuScenes, whose terms of use are non-commercial. Read them before use.
2. No declared checkpoint license. OpenMMLab publishes this file from its model zoo without any per-object license. The mmdetection3d repository states only that "this project is released under the Apache 2.0 license", which covers the code. LibreYOLO's maintainer approved this mirror by treating that project-level Apache-2.0 statement as the redistribution basis. This is LibreYOLO's disclosed interpretation, not a clarification from the OpenMMLab authors.

The FCOS3D architecture and LibreYOLO's code are unrestricted, as is any model you train from scratch. Only these pretrained weights are limited.

Source

• Publisher: open-mmlab/mmdetection3d model zoo
• Original URL: fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645-8d806dc2.pth
• SHA-256: 8d806dc2ecae85bc8eaba1f16dccdf03459317ca6aed7984cfda33c2a2bc33a8
• Size: 220,360,589 bytes
• Published July 2021.

This is the finetuned variant (depth weight 1.0), not the 1x base model. Published nuScenes val results: 32.1 mAP, 39.5 NDS.

The LibreYOLO implementation that consumes it was adapted from mmdetection3d at fe25f7a and mmdetection at cfd5d3a, both Apache-2.0.

Modifications

None. The checkpoint is mirrored byte-for-byte, retaining its original filename, state_dict/meta structure and serialization. No tensor values or metadata were changed. It is not converted to LibreYOLO's v1.0 checkpoint schema — the adapter loads the unchanged official format directly.

Classes

The ten nuScenes detection classes, in the official order. LibreYOLO validates the checkpoint's meta.CLASSES against that order and refuses a mismatch.

LibreYOLO usage

import numpy as np
from libreyolo import LibreFCOS3D

model = LibreFCOS3D("fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645-8d806dc2.pth")
result = model.predict("image.jpg", intrinsics=np.load("intrinsics.npy"))
print(result.boxes3d.xyz)

FCOS3D is a sibling API, not part of the generic LibreYOLO(...) factory, because prediction requires original-image pinhole camera intrinsics. Download this file and pass its local path. Inference is CPU or CUDA; MPS is unsupported (torchvision deform_conv2d). Training, validation, tracking and export are not implemented.

Validation

LibreYOLO's native implementation was verified against the pinned upstream Python implementation using this checkpoint: every state_dict key and shape loads strictly, and final CPU detections match across six calibrated nuScenes camera views at confidence thresholds 0.05, 0.15 and 0.30. This is adapter parity, not an independent 3D accuracy benchmark; use the official nuScenes evaluator for accuracy.

License

See LICENSE for the upstream Apache-2.0 code license and the checkpoint's separate terms, and NOTICE for provenance and mirroring details.