WALDO40 YOLOv26 P2
Freshly trained P2 object detectors for twelve-class aerial imagery. These
checkpoints come from the independent yolo-drone-v26b PyTorch implementation
and use Obnubilated environmental and camera augmentations.
N and S are available now. M and L are pending. Their existing training queue was intentionally paused while throughput improvements are integrated; no incomplete M checkpoint is published.
Available checkpoints
Both released models completed the full 300-epoch, 640 px training schedule. The published file is the best checkpoint selected by validation AP-small, so its checkpoint epoch can precede the final training epoch.
| Size | File | Best checkpoint epoch | mAP50-95 | mAP50 | AP-small | SHA-256 |
|---|---|---|---|---|---|---|
| N | weights/waldo40-yolov26n-p2-obnubilated.pt |
256 | 0.4237 | 0.6318 | 0.3601 | c0ecd043158f04e32a3f4c9ccc739c040971a9e7d29af5b7865e84f900ec1ce1 |
| S | weights/waldo40-yolov26s-p2-obnubilated.pt |
217 | 0.4596 | 0.6760 | 0.4028 | 63a3f42470dc61fa47baf0a626e43d733cb74e361e82fd58325dcb8f5549e8bf |
| M | — | — | Pending | — | — | Training paused for throughput work |
| L | — | — | Pending | — | — | Not started |
The metrics are single-seed engineering results on the frozen 710-image WALDO 4.0 clean validation split. They are not multi-seed uncertainty estimates.
The downloadable files are inference-only packages. Optimizer, scheduler,
gradient-scaler, provider, and private host state were removed. The model and
EMA tensors are byte-for-byte equal to the terminal training artifacts. The
mapping from terminal artifact hashes to released hashes is recorded in
release-manifest.json.
WALDOv3.0, YOLOv8, and YOLOv26 comparison
The WALDO 4.0 columns use the same 640 px class contract and clean validation split. The YOLOv26 lane uses Obnubilated training only.
The WALDOv3.0 values are native validation metrics embedded in the public
640 px P2 checkpoints at release revision 6ad69ea. They are historical
context measured on different validation data, so they must not be subtracted
and presented as a controlled improvement.
| Size | WALDOv3.0 P2 native-val mAP50-95 | WALDO 4.0 YOLOv8 Standard mAP50-95 | WALDO 4.0 YOLOv8 Obnubilated mAP50-95 | WALDO 4.0 YOLOv26 Obnubilated mAP50-95 | Status |
|---|---|---|---|---|---|
| N | 0.3865 | 0.4425 | 0.4377 | 0.4237 | Released |
| S | Not available* | 0.4978 | 0.5002 | 0.4596 | Released |
| M | 0.5161 | 0.5312 | 0.5451 | Pending | Training paused |
| L | 0.5385 | 0.5438 | 0.5529 | Pending | Not started |
* The public WALDOv3.0 release contains 640 px P2 checkpoints for N, M, and L, but no S checkpoint. Previously posted same-split WALDOv3.0-on-WALDO4.0 figures were withdrawn pending investigation and are not used here.
The released YOLOv8 checkpoints are in the public WALDO40 YOLOv8 repository.
Classes
The zero-based class order is:
| ID | Class | ID | Class |
|---|---|---|---|
| 0 | LightVehicle |
6 | Container |
| 1 | Person |
7 | Truck |
| 2 | Building |
8 | Gastank |
| 3 | UPole |
9 | Digger |
| 4 | Boat |
10 | Solarpanels |
| 5 | Bike |
11 | Bus |
Inference
inference.py downloads the selected config and checkpoint, verifies its
SHA-256, and invokes the clean-room v26b-infer command:
python -m pip install "huggingface_hub>=1,<2"
python inference.py image.jpg --size s --device cpu --output runs/predict
The clean-room runtime repository remains private during its final source
release review. Consequently the public weights and configuration are
available now, but the command above additionally requires an authorized local
installation that provides v26b-infer. The card will gain a public install
command when that runtime is released; no unrelated implementation is implied
to be compatible.
License
The released checkpoints, configurations, and inference helper are available under the MIT License. No converted or third-party pretrained detector weights were used by this lane.