Model packages and JSON contracts

ZIP structure

Place model_info.json and coco_categories.json at the archive root. Weights are addressed by the relative weights_path. Include additional files, including model.py, when required by the selected custom-model protocol; a Python implementation is not mandatory for every built-in type.

model-package.zip
├── model_info.json
├── coco_categories.json
└── weights/
    └── model.onnx

Model metadata and profiles

model_type must be registered in the target runtime and task_type must match the model. A custom type requires task_type, deploy_profile, img_size and weights_path. A weights-file extension alone does not prove runtime support.

Example model_info.json

Replace angle-bracket placeholders with values supported by your runtime. The 640 × 640 input is illustrative; use the actual model dimensions.

{
  "model_type": "<supported-model-type>",
  "task_type": "<supported-task-type>",
  "img_size": {
    "height": 640,
    "width": 640
  },
  "weights_path": "weights/model.onnx",
  "deploy_profile": {
    "default_profile": "<runtime-profile>",
    "profiles": [
      "<runtime-profile>"
    ]
  }
}

Category contract

The root object contains categories. Each id must match a class produced by the model. Threshold and colour support processing and visualisation; is_violation does not replace event-creation logic. The legacy schema keys ru_name and full_ru_name are retained for compatibility; their example values are English display text.

{
  "categories": [
    {
      "id": 1,
      "name": "example_object",
      "ru_name": "Object",
      "full_ru_name": "Example object",
      "supercategory": "object",
      "threshold": 0.5,
      "color": [
        0,
        255,
        0
      ],
      "is_violation": false,
      "is_primary": true,
      "exclude": false
    }
  ]
}

Pre-upload validation

• Parse both JSON files and validate field types.
• Verify every relative weights path and required protocol file.
• Match category IDs to actual model outputs.
• Validate each declared profile on the target runtime.
• Wait for automated checks after upload, then run the model on a representative asset.

tested establishes technical operation, not accuracy on a site’s cameras.