AoE2 entity detector
This model detects 60 entity classes in Age of Empires II: Definitive Edition
screenshots. model.onnx contains the weights used by this project's Mac-hosted
detector server. The model repository is an artifact registry, not a hosted
inference API. Future versions replace model.onnx in a new commit;
consumers must pin a commit or release tag.
Intended use
The supported use is screen-based perception for the project's game agent. The model was trained on synthetic scenes generated from game sprites and annotated real-game screenshots. Neither the training images nor extracted sprites are included in this artifact release.
The model is fine-tuned from the Ultralytics YOLO26n base model and exported to ONNX. Its metadata identifies the Ultralytics AGPL-3.0 license.
License and terms
The model weights are available under the GNU Affero General Public License v3.0 (AGPL-3.0). You may use, modify, and redistribute them, including commercially, subject to that license. This project imposes no additional noncommercial-use restriction or fee.
Users integrating the model into an application or service must meet applicable AGPL-3.0 obligations, including corresponding-source requirements. For different terms covering Ultralytics technology, contact Ultralytics about its Enterprise license. This is an independent fine-tune, not an official Ultralytics release. The model is provided as-is, without warranty.
Inference contract
See inference-contract.json and classes.yaml. Input is a 1280 × 1280 RGB
letterboxed image, normalized to float32 values in [0, 1] and transposed to
NCHW. The model returns up to 300 rows of
[x1, y1, x2, y2, confidence, class_id] in model-input pixel coordinates.
The detector server applies class-specific confidence thresholds and maps boxes
back to the original screenshot; its client handles tracking and deduplication.
Loading the ONNX file alone does not reproduce those application-level results.
Evaluation and limitations
On the project's 32-image real-frame validation split (1,212 labeled objects), the served ONNX checkpoint reaches micro precision 0.753, recall 0.633, and F1 0.688 at IoU ≥ 0.50. This uses the detector server's single-pass 1280-pixel preprocessing, its deployed per-class confidence thresholds, and the client's classwise NMS. The evaluation used ONNX Runtime's CPU provider and excludes temporal tracking, cached detections, and gameplay.
This is validation, not an independent test benchmark. Training and validation images came from the same January 2026 capture sessions and can contain nearby frames; the split was by image rather than game/session. The existing confidence thresholds were also developed using the project's validation workflow. The labels and screenshots are not released with the model. This overall score is dominated by common classes and does not establish reliable detection of rare objects. The full-screen 1280 × 831 validation images also do not establish performance on other game captures, UI scales, maps, or versions. The model's ability to support reliable gathering or combat has not been validated by this score.
See the complete evaluation protocol and per-class results and machine-readable counts and hashes. A future independent, session-held-out test should be reported separately rather than replacing this validation result.
Provenance
- Served registry artifact:
model.onnx - Model SHA-256:
515a018bc2190fdf5427a01ff21e294331324929c8603d870c18255626cee8fd - Class schema SHA-256:
5365dcf538d16f9b237070a5e9c7609028314794dd8e544233eecb76a09de717 - Detector source revision:
ca4b06310a49d46381c043bec3a35618c0e5dee2
For reproducible game runs, download a full-commit-pinned revision and verify both checksums before starting the detector server. Do not download weights in the agent's frame-processing loop.
- Downloads last month
- -