Models and Resources

This document describes the model and resource organization that can be confirmed in the current repository.

Resource Directories

Directory	Purpose
data/resource/aiboxresource_bm1688	Resources for the Sophon BM1688 release package.
data/resource/aiboxresource_cv186x	Resources for the Sophon CV186X release package.
data/resource/aiboxresource_x86	Resources for the x86 Docker / CPU backend.
data/resource/aiboxresource_rknn	Rockchip target resources and model overlays, combined with base templates

The resource directory is selected through RESOURCE_DIR at build time.

RKNN builds reuse base templates from aiboxresource_x86 and apply target-specific resources. RK3576 uses aiboxresource_rknn; RV1126B generates output/platform-artifacts/rv1126b/resource-overlay from its platform profile and artifact manifest. Base templates, target overlays and the model directory determine the packaged resources. Refer to the package manifest for its actual contents.

Model Templates

A template defines configuration and parsing options. It does not mean the weights are included or the model is supported on every platform; availability depends on the backend, target chip, runtime and installed resources.

Model templates are located at:

data/resource/*/model_template

The templates currently visible include:

• YOLO detection templates
• YOLO classification templates
• DINO
• SAM2
• feature
• keypoints
• Qwen3 / Qwen3VL
• OCR

License-Plate OCR Package

OCR recognizes plates already located by landmarks; it does not perform full-image text detection. Install the model package through device resources. This repository provides only the ocr.json template and never includes weights or private dictionaries.

Each external OCR package must contain:

• config.json with "model_type": "ocr";
• the current-platform model file (.nn for Sophon, .onnx for x86);
• the .txt character table named by its configuration.

models[0].params must explicitly bind the table and CTC mapping: character_table_file must name a root-level .txt file, ctc_blank_index identifies the blank class, ctc_prepend_tokens and ctc_append_tokens define additional leading and trailing model classes, and ctc_class_count must equal the model output's final dimension. The table plus leading and trailing tokens must match CTC indices exactly; a mismatch fails OCR initialization or inference instead of silently dropping characters. The Add Model wizard requires both the OCR model and its character table and generates these fields.

Legacy license-plate model 2000007 uses 6,625 classes with a 6,624-line table: entry zero is blank and the configuration appends one ASCII space for the final class. Decoding removes blank and consecutive duplicate classes.

A PP-OCR Chinese recognition model may use the standard 6,623-entry table without blank. For a 6,625-class output, the system explicitly prepends blank and appends an ASCII space.

Connect the scene nodes as “plate detection/association → four-point plate landmark → text recognition → event report”. OCR accepts only a strict four-point plate quadrilateral in top-left, bottom-left, bottom-right, top-right order; empty recognition results do not produce an alarm. In the event-report node, select the Vehicle Property alarm attribute to expose plateSrc, plate, and the rectified plate crop in the event.

Layout and Components

Resource layout files include:

data/resource/*/layout/modelComponents.json
data/resource/*/layout/actions.json
data/resource/*/layout/linkageStorages.json

These files affect the frontend configuration items, model component parameters, action nodes, and linkage strategies.

Algorithm Templates

Algorithm templates are located at:

data/resource/*/algorithm_template

Templates for vision-language models, DINO, YOLO, and related models are visible here. For an official release, it is recommended to generate a "template catalog table" from the current resource directory to avoid a hand-written list going stale.

x86 and Sophon Differences

x86 path:

data/resource/aiboxresource_x86

Sophon path:

data/resource/aiboxresource_bm1688
data/resource/aiboxresource_cv186x

The code shows the differences between handling x86 ONNX files and Sophon model packages. The full model porting workflow should be re-validated against the currently releasable model packages.

Resource Licensing Notes

The resource tree contains model, algorithm, and layout templates as well as selected public example weights and chip-converted artifacts. The repository's Apache-2.0 source-code license does not automatically cover model files; use the bundle manifest and directory-level license as the authority:

• Prebuilt components in the prebuild/ directory require separate distribution-license review.
• Ultralytics YOLOv8 chip artifacts under data/resource/aiboxresource_rknn/model-artifacts/ are AGPL-3.0 community examples, not commercial model deliverables.
• Commercial or proprietary models use independent source, weight, training, and license records and do not reuse the community example bundle identity.
• Whether the model encryption feature is included in the current build depends on the CMake option COSMO_MODEL_GUARD.
• If you introduce models from a third-party model ecosystem, follow the license requirements of the corresponding models.