LibreWildDet3D

Mirror of the WildDet3D Stage 3 paper checkpoint for use with LibreYOLO's optional LibreWildDet3D adapter.

License and use restrictions

This checkpoint is not covered by LibreYOLO's MIT license. It is provided under the custom SAM License, which includes field-of-use and trade-control restrictions. Read and comply with the complete license before downloading or using the checkpoint.

Source

• Publisher: allenai/WildDet3D
• Upstream repository revision: 1b8aa52
• Upstream Hugging Face revision: 161edb26239f58d8973045517d9c8b179c515f73
• Original filename: wilddet3d_alldata_all_prompt_v1.0.pt
• SHA-256: f8b6a9e548f733ba62625a0d2adc4b0f4fdb6007ee11d9927f9c1027010fee57

The checkpoint is mirrored byte-for-byte. It contains the model's Stage 3 weights and retains its original filename and format. No tensor values, metadata, or serialization structure were changed.

Classes and prompts

WildDet3D has no fixed class list. It accepts text vocabularies such as ["car", "person"], as well as box and point prompts. Returned class ids index the supplied text vocabulary or prompt order.

LibreYOLO usage

import numpy as np
from libreyolo import LibreWildDet3D

model = LibreWildDet3D()
result = model.predict(
    "image.jpg",
    intrinsics=np.load("intrinsics.npy"),
    text=["car", "person"],
)
print(result.boxes3d.xyz)

The separately licensed WildDet3D runtime must be installed. On macOS, LibreYOLO uses an isolated CPU worker; see the LibreYOLO detect3d contract for the tested setup. MPS is not supported by the upstream image path. Omitting model_path downloads this revision-pinned mirror and verifies its SHA-256; an explicit local upstream checkpoint remains supported.

Validation

Using ten identical images, prompts, calibrations, thresholds, and this checkpoint:

• LibreYOLO CUDA matched direct upstream CUDA exactly for all returned numeric fields (max_abs_diff == 0.0).
• LibreYOLO macOS CPU selected the same detections and classes in all ten cases. Maximum CPU/CUDA drift was 0.173 pixels for 2D boxes and 0.0374 metres for 3D centers.

This verifies adapter parity. It is not a 3D accuracy benchmark; the comparison images did not provide calibrated camera intrinsics or 3D ground truth.

License

See LICENSE for the complete SAM License and NOTICE for provenance and mirroring details.