MirilAI/WALDO40-YOLOv8

WALDO40 YOLOv8 P2

YOLOv8 P2 detectors trained at 640 px on WALDO 4.0 with the latest cleaned labels. Every model size is provided in two directly comparable variants:

• Standard uses the conventional color augmentation stack.
• Obnubilated replaces that color jitter with randomized, geometry-aware relighting, depth-varying fog, and camera/image degradation. Obnubilated is sampled on half of training inputs; the ordinary spatial transforms remain.

This repository contains all eight best-validation checkpoints:

Size	Training variant	Checkpoint	Validation mAP50-95	Validation mAP50
N	Standard	weights/waldo40-yolov8n-p2-standard.pt	0.4425	0.6520
N	Obnubilated	weights/waldo40-yolov8n-p2-obnubilated.pt	0.4377	0.6501
S	Standard	weights/waldo40-yolov8s-p2-standard.pt	0.4978	0.7191
S	Obnubilated	weights/waldo40-yolov8s-p2-obnubilated.pt	0.5002	0.7225
M	Standard	weights/waldo40-yolov8m-p2-standard.pt	0.5312	0.7525
M	Obnubilated	weights/waldo40-yolov8m-p2-obnubilated.pt	0.5451	0.7684
L	Standard	weights/waldo40-yolov8l-p2-standard.pt	0.5438	0.7677
L	Obnubilated	weights/waldo40-yolov8l-p2-obnubilated.pt	0.5529	0.7800

The values above come from the best checkpoint of one deterministic seed and a clean validation split. They are engineering results, not a multi-seed scientific claim. The frozen release dataset declares training and validation splits only; no current test result is claimed.

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. Its values remain pending until each full training run and terminal evaluation complete.

The WALDOv3.0 values are the native validation metrics embedded in the public 640 px P2 checkpoints at release revision 6ad69ea. They are useful historical context, but they were measured on the original WALDOv3.0 validation data—not the WALDO 4.0 clean split. Dataset composition and the split changed, so these columns 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	Pending	Awaiting YOLOv26 terminal evaluation
S	Not available*	0.4978	0.5002	Pending	Awaiting YOLOv26 terminal evaluation
M	0.5161	0.5312	0.5451	Pending	Awaiting YOLOv26 terminal evaluation
L	0.5385	0.5438	0.5529	Pending	YOLOv8 pair complete; awaiting YOLOv26

* The public WALDOv3.0 release contains 640 px P2 checkpoints for N, M, and L, but no S checkpoint. The previously posted same-split WALDOv3.0-on-WALDO4.0 figures have been withdrawn pending investigation and are not used in this comparison.

The public WALDO40 YOLOv26 repository will receive checkpoints, hashes, and measured values only after validation.

What Obnubilated changes

These are representative draws from one source image. They demonstrate the type of variation used during training, not a fixed filter applied at inference time.

Source	Geometry-aware relighting	Depth-varying fog

Classes

The class indices are:

1. LightVehicle
2. Person
3. Building
4. UPole
5. Boat
6. Bike
7. Container
8. Truck
9. Gastank
10. Digger
11. Solarpanels
12. Bus

Inference

Install the pinned inference dependencies:

python -m pip install -r requirements.txt
python inference.py image.jpg --size m --variant obnubilated --output runs/predict

Training configuration

• 300 epochs
• 640 px inputs
• P2 detection head for small aerial objects
• SGD optimizer
• Obnubilated environment and post-3D augmentation, sampled with probability 0.50 in the Obnubilated variant
• Standard and Obnubilated variants retain the same P2 architecture, training budget, spatial augmentation, split, and seed; the environment/color path is the intended comparison variable
• Python 3.11 and PyTorch 2.6.0

License

The model checkpoints in this repository are distributed under the MIT License. See LICENSE for the full terms.