Overview

The Generic Classifier node is designed to retrieve classifier results by analyzing the similarity of text prompts and the images of objects within a Region of Interest (ROI), or the contents of a ROI image. This functionality is useful for applications requiring detailed classification of objects or areas within a video feed.

Inputs & Outputs

• Inputs: 1, Media Format: Raw Video
• Outputs: 1, Media Format: Raw Video
• Output Metadata: nodes.node_id, recognized_objs, recognized_obj_ids, recognized_obj_count, recognized_obj_delta, value_changed_delta, unrecognized_obj_count, unrecognized_obj_delta

Pipeline setup

Choose the processing mode based on what should be classified:

• ROIs, at an interval classifies each configured ROI periodically. Use this for scene state such as an occupied bay, an open door, or a blocked aisle.
• ROIs, upon a trigger classifies an ROI only when the trigger becomes true. Place the node after the metadata-producing node referenced by the trigger.
• Objects in an ROI classifies crops from objects already present in frame metadata. Place an object detector—and normally a tracker—upstream. Stable object IDs allow the lookup limits, size-change checks, and continuous mode to work as intended.

Use the resulting ROI labels or object attributes in downstream Filter, Event, notification, or storage nodes. Start with display_objinfo and debug enabled while tuning, then disable them after validating the labels and thresholds on representative footage.

Properties

Guidelines

Model and prompt guidance

The classifier uses a SigLIP-based model to compare each visual description with the image. Similarity scores are evaluated independently for each description, so:

• Write each class as a concrete visual description of what the crop should look like—not as an instruction, keyword, metadata label, or abstract yes/no question.
• Describe competing classes at a similar level of detail and from the expected camera viewpoint.
• Prefer visible positive evidence (clothing, color, object, posture, scene state) over pure negation such as “not safe.”
• In multi mode, several descriptions can independently exceed their thresholds; the scores do not form a probability distribution that sums to one.
• Tune thresholds separately for each model option. Record scores with debug, begin permissively, then raise each threshold until false positives are rejected. A per-class @threshold override is useful when one visual concept is consistently easier or harder than the others.

Each non-empty line in input_text has this form:

<visual description>=<output label>[@threshold]

The output label and threshold are optional. The model only receives the description to the left of =; the label is what the node writes to metadata.

Prompt examples

For a first deployment, test prompts against daytime, nighttime, near/far, partially occluded, and empty-ROI examples. Include a fallback label only when downstream logic needs an explicit value for “none of the configured descriptions matched.”

Metadata

The output metadata includes detailed information about the classification results, such as the types of objects recognized, their counts, and any changes in their recognition status. This allows for a comprehensive understanding of the scene being analyzed.

Output Metadata

The fields below are declared by this node's metadata schema; the JSON values are representative examples.

JSON example

{
  "nodes": {
    "generic_classifier1": {
      "label_changed_obj_delta": 0,
      "objects_of_interest_keys": [],
      "recognized_obj_count": 0,
      "recognized_obj_delta": 0,
      "recognized_obj_ids": [],
      "rois": {
        "roi1": {
          "label": "example",
          "label_available": false,
          "label_changed_delta": false
        }
      },
      "type": "generic_classifier",
      "unrecognized_obj_count": 0,
      "unrecognized_obj_delta": 0
    }
  }
}

Object labels and attributes

• Object labels/classes added: In ROI mode, the configured ROI label is added as an object with class 10800.
• Object attribute labels/classes added: ROI objects receive generic_classifier_roi (10800); matched configured output labels, or the configured fallback label, use class 10801 on ROI or upstream objects.