fireviewer/dinov3-pointing-pilot-v1

FireViewer DINOv3 pointing pilot v1

Research archive, currently publicly visible on the Hub, of a five-epoch DINOv3 ViT-B/16 multi-task pilot for visible fire/smoke presence, segmentation, point localization and explicit visual abstention.

This repository preserves a completed, reloadable experiment. It is not a production model and not independently benchmarked. The retained release records state that public release was not approved; current public visibility does not resolve that restriction or establish redistribution permission. The word “pilot” is part of the model identity and should not be removed in downstream documentation.

Model contract

• Base model: facebook/dinov3-vitb16-pretrain-lvd1689m
• Immutable base revision: 5931719e67bbdb9737e363e781fb0c67687896bc
• Architecture: DinoV3MultiTaskModel, schema 4
• Input size: 448 × 448 RGB
• Heads: segmentation, point heatmap, visual abstention, fire/smoke presence
• Presence labels: flame_visible, smoke_visible
• Training: full-model fine-tuning, five epochs, 8,192 balanced samples/epoch
• Selection: minimum validation loss, epoch 5

Validation record

The metrics below come from the campaign validation split. They are not an independent benchmark and must not be compared directly with the frozen object detection benchmark used for YOLO, D-FINE and RT-DETR. RF-DETR has not yet passed that independent protocol.

Metric	Epoch 5 result	Evidence boundary
Validation loss	0.02892	11,007 validation rows
Segmentation IoU	0.86718	Aggregate is strongly influenced by negative/background rows
Segmentation Dice	0.92887	Same boundary as IoU
Point PCK@10	0.55556	Only 9 positive point rows
Point-negative specificity	0.99402	6,188 negative point rows
Presence exact-match accuracy	0.91778	11,007 presence rows
Flame-presence F1	0.87507	Validation split only
Smoke-presence F1	0.93582	Validation split only

The training split contains 36,039 rows but only 134 rows in the positive pointing role. This imbalance and the nine positive validation points make the localization score too uncertain for a scientific or operational claim. See validation-summary.json and the machine reports in reports/.

Artifacts

• model.safetensors: preferred inference weights; 271 tensors and 94,127,621 parameter values were converted and byte-for-byte tensor checked.
• checkpoints/best.pt: native epoch-5 selected checkpoint.
• checkpoints/last.pt: full optimizer/scheduler resume checkpoint at epoch 5.
• dinov3_adapter.py: exact FireViewer schema-4 architecture used for loading.
• config.json: pinned architecture, base revision and artifact paths.
• artifact-manifest.json: release inventory and recorded checkpoint identity.

Loading the safe weights

Access to the gated DINOv3 base configuration and acceptance of Meta's DINOv3 terms are required. The published safetensors file contains the full trained backbone and all FireViewer heads; the base repository is used to instantiate the pinned architecture configuration.

from dinov3_adapter import DinoV3MultiTaskModel

model = DinoV3MultiTaskModel.from_safetensors(
    "model.safetensors",
    model_id="facebook/dinov3-vitb16-pretrain-lvd1689m",
    revision="5931719e67bbdb9737e363e781fb0c67687896bc",
    image_size=448,
    token=True,
)
network = model.network.eval()

The caller remains responsible for applying the same image normalization and for interpreting logits from each head. Do not treat a heatmap maximum as a verified geographic coordinate.

Rights, limitations and safety

The derived weights remain subject to the DINOv3 terms and to applicable source-specific obligations documented by the training receipts. Public or private access does not grant redistribution rights. See RIGHTS_AND_ATTRIBUTION.md.

This model is not a certified warning system. It must not be used alone to confirm a fire, trigger an alert, order an evacuation or direct emergency response.

Visibility and retained release policy

The Hub metadata reports this repository as public on 19 September 2026. The retained artifact-manifest.json and validation-summary.json still describe its historical private pilot/release policy. Those evidence files are preserved. This card corrects the description of observed visibility; it does not approve a public release, change repository access, clear source rights or promote the pilot. A maintainer decision on publication rights remains outstanding.

Sensitive outputs require human review. A heatmap point is an image-space proposal, not a verified base in the terrain or a hidden ignition origin. The nine positive validation points cannot support a robust localization claim. No independent positive-point benchmark or new evaluation is supplied by this documentation update.

Documentation, revision and stewardship

Documentation reviewed on 19 September 2026, using the FireViewer documentation pack and the repository at ea193c71e8034c2477596d3648d727ceb9e78a03. This is a documentation and metadata check, not a new model evaluation, training run, data acquisition or full artifact verification. Historical receipts keep their original dates and identifiers; manifest hashes of earlier README versions describe those earlier releases.

The association FIRE-VIEWER provides administrative and financial stewardship for resources under its control. Technical governance is maintainer-led. Pre-association rights, UWD components, upstream licences and source-specific rights remain separate; repository placement does not transfer ownership or grant new rights.

Canonical documentation · Status vocabulary · Institutional website. Contact: contact@fire-viewer.fr.