| --- |
| license: cc-by-4.0 |
| library_name: onnxruntime |
| pipeline_tag: image-classification |
| tags: |
| - mobile-legends |
| - draft-pick |
| - obs |
| - computer-vision |
| - onnx |
| - mobilenet-v3 |
| --- |
| |
| # SHAYYZ MLBB Draft Classifier |
|
|
| SHAYYZ MLBB Draft Classifier is a lightweight visual classifier for recognizing |
| hero portraits in Mobile Legends: Bang Bang draft slots. It is designed for the |
| local SHAYYZ MLBB OVERLAY OBS workflow and does not use a game API, inspect game |
| memory or traffic, or automate game input. |
|
|
| ## Model details |
|
|
| - Architecture: MobileNetV3-Small |
| - Revision: `87f29aa7f9e0` |
| - Classes: 133 heroes plus `empty` and `unknown` |
| - Input: RGB `float32`, NCHW `1 x 3 x 224 x 224` |
| - Normalization: ImageNet mean `[0.485, 0.456, 0.406]` and standard deviation |
| `[0.229, 0.224, 0.225]` |
| - Output: 135 unnormalized logits in the order recorded by `manifest.json` |
| - Runtime: ONNX Runtime CPU |
| - Precision: FP32 |
|
|
| The model was initialized from the official torchvision |
| `MobileNet_V3_Small_Weights.DEFAULT` checkpoint and fine-tuned locally. |
|
|
| ## Evaluation |
|
|
| | Evaluation | Top-1 accuracy | Macro recall | Unknown false-accept rate | |
| | --- | ---: | ---: | ---: | |
| | Synthetic validation, 270 images | 100% | 100% | — | |
| | Synthetic test, 270 images | 100% | 100% | 0% | |
|
|
| The exported ONNX logits differed from the PyTorch checkpoint by at most |
| `0.0003394`. The production Bun/ONNX Runtime path classified all 270 synthetic |
| test images correctly at a mean of 8.4 ms per image on an Apple M3 Pro CPU. |
|
|
| These results measure a deterministic synthetic holdout derived from the same |
| private portrait source family used for training. They do not establish live |
| game accuracy or robustness to future game UI and artwork updates. Keep the |
| overlay in proposal mode until a separate live replay benchmark validates the |
| current game build and capture profile. |
|
|
| ## Intended use |
|
|
| The model may propose hero selections from already-visible draft-slot crops for |
| a human-operated broadcast overlay. It is not intended for player automation, |
| gameplay decisions, identity recognition, surveillance, or unattended control |
| of MLBB or OBS. |
|
|
| ## Training data |
|
|
| Training used 3,240 private 224 x 224 synthetic samples: 24 variants for each of |
| 133 heroes plus balanced `empty` and `unknown` classes. Augmentations reproduce |
| face-focused framing, transparent edge fades, team tinting, ban dimming, blur, |
| and compression. The private dataset and original game artwork are not included |
| in this repository or model release. |
|
|
| ## License and attribution |
|
|
| The model weights and this model card are released under CC BY 4.0. Attribute |
| them to Aung Min Khant and name **SHAYYZ MLBB Draft Classifier**. |
|
|
| Mobile Legends: Bang Bang and its hero artwork are owned by Moonton. Those game |
| assets are not distributed here and are not covered by this license. This model |
| is an independent, unofficial project and is not endorsed by or affiliated with |
| Moonton. |
|
|
