Research at Noble

Architecture choices

NobleAIM is built on the YOLO (You Only Look Once) family of single-pass object detectors, originally introduced by Redmon et al. and now maintained as the YOLO11 and YOLO26 generations by Ultralytics. We hold an official Ultralytics enterprise licence, which means every model we ship is properly licensed for commercial use and benefits from upstream architectural improvements.

Three task heads are in active use across our supported titles: detection (axis-aligned bounding boxes), pose estimation (skeletal keypoints, building on the line of work from Cao et al. 2017 and Newell et al. 2016), and instance segmentation (pixel-perfect masks, descended from He et al. 2017). Each task head is selected per game based on what the game’s visual signal actually rewards — pose for unobstructed enemies in cover-heavy titles, segmentation where bounding boxes drift under heavy occlusion, detection where raw inference speed matters most.

We deliberately prefer many small per-game models over a single generalist. A specialist YOLO11n trained on a single game outperforms a YOLO11m trained on a mix of games for any given title, at a fraction of the inference cost — a result consistent with the broader pattern in fine-grained recognition research. The Noble Marketplace exists in part to make this per-game specialisation distributable.

Training methodology

Datasets are curated per game from live gameplay footage, with manual annotation passes by domain-specialist labellers familiar with each title’s specific visual vocabulary (operator silhouettes, ability VFX, skin variants, weather and time-of-day variation). We invest in real-gameplay footage over synthetic data: a model trained on actual capture-card output generalises substantially better to actual capture-card output than one trained on rendered frames, even where the renderer is the same game engine.

Standard YOLO augmentation (mosaic, mixup, HSV jitter, random affine) is applied with per-game tuning. Skin and cosmetic robustness is treated as a first-class concern — for titles where character cosmetics are frequently refreshed (Apex Legends, Fortnite, Marvel Rivals), augmentation pipelines explicitly include cosmetic-variant sampling so detection survives content updates without immediate model retraining.

Training is performed on Ultralytics Hub, which Noble recommends to anyone building custom models for use with NobleAIM. Hub is maintained by the team behind YOLO itself, supports the same YOLO11 and YOLO26 detect, pose, and segment task heads NobleAIM uses at runtime, and exports straight to ONNX. NobleAIM’s built-in Engine Builder then converts those ONNX files into TensorRT, ONNX-CUDA, DirectML, or OpenVINO runtime engines targeted at the user’s GPU. The same tooling Noble uses internally is what external model authors use.

Inference backends

NobleAIM ships four inference backends and selects the strongest one for the user’s hardware automatically. This is unusual for tooling in this space, where single-backend (typically NVIDIA-only) is the norm. We treat the inference path as a research-relevant variable, not an implementation detail.

Backend selection is cached per session at startup using a hardware probe; per-frame backend overhead is zero. Internal head-to-head measurements across the four backends inform our default-selection heuristics; cross-backend results will be published once the methodology described below stabilises.

Real-world evidence

Qualitative gameplay evidence is published at /gameplay/: live capture of NobleAIM operating across multiple titles, encoded for the website at native capture rate so detection behaviour is visible frame-by-frame on inspection. These clips are deliberately not staged — the cuts shown are continuous gameplay segments.

We treat these clips as qualitative evidence, not as benchmarks. They demonstrate the system working under real conditions; they do not yet quantify it.

What we deliberately do not yet claim

This space — external AI aim assist — has a strong incentive to publish unsubstantiated marketing numbers. We have chosen not to.

• We do not publish frames-per-second figures without identifying the exact hardware, driver, capture source, and model version they were produced on.
• We do not publish detection accuracy (mAP, precision, recall) without a frozen, held-out test set distinct from training data.
• We do not publish aim-improvement metrics that rely on subjective player evaluation.
• We do not publish head-to-head comparisons against competing systems on data either system was likely trained on.

Per-game held-out test sets, a reproducible benchmark runner across the four inference backends, and a multi-GPU test bench are in active internal development. Methodology will be documented before numbers are published; numbers without methodology are noise.

Research foundations

NobleAIM is engineering on top of an active research lineage. The papers and projects below directly inform the architecture choices and training methodology described above.

• Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). arXiv:1506.02640
• Redmon, J., & Farhadi, A. (2018). YOLOv3: An Incremental Improvement. arXiv:1804.02767.
• Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv:2004.10934.
• Wang, C.-Y., Bochkovskiy, A., & Liao, H.-Y. M. (2023). YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. CVPR 2023. arXiv:2207.02696
• Jocher, G., Chaurasia, A., & Qiu, J. (2023). YOLO by Ultralytics (software and model family). github.com/ultralytics/ultralytics
• He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. IEEE International Conference on Computer Vision (ICCV). arXiv:1703.06870
• Cao, Z., Simon, T., Wei, S.-E., & Sheikh, Y. (2017). Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. CVPR 2017. arXiv:1611.08050
• Newell, A., Yang, K., & Deng, J. (2016). Stacked Hourglass Networks for Human Pose Estimation. European Conference on Computer Vision (ECCV). arXiv:1603.06937
• Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014). Microsoft COCO: Common Objects in Context. ECCV 2014. arXiv:1405.0312
• NVIDIA Corporation. TensorRT Developer Guide. docs.nvidia.com/deeplearning/tensorrt
• Microsoft. Direct Machine Learning (DirectML). learn.microsoft.com/en-us/windows/ai/directml
• Intel Corporation. OpenVINO Toolkit. docs.openvino.ai