Feature Vector Clustering

The Feature Vector Clustering feature creates clusters of input feature vectors. The input vectors belonging to the same cluster are associated based on similarity. Recognized feature vectors used as input for creating a cluster are called feature vector observations.

All clusters are stored in the same database regardless to which stream the input feature vectors come from. Each cluster is represented by a single feature vector which can be different from all feature vector observations.

The feature implements cluster updates which can be the following:

• Updating an existing cluster based on an observed new feature vector that is similar to the representative feature vector of the concerning cluster stored in the data base
• Creating a new cluster based on an observed new feature vector that is not similar to any existing representative feature vectors.

Creating a new cluster is called cluster realization. Realizing a cluster and handling newly relaized clusters depend on Feature Vector Clustering configuration parameters defined in Configuring Feature Vector Clustering.

• Merging two or more clusters into one when they are more similar to each other than a configurable threshold. See Feature Vector Cluster Merging.

A typical use case of the feature is pre-clustering for Reidentification. The Reidentification feature matches observed feature vectors to the representative feature vectors of relized clusters that are stored in a database. Single feature vectors can be unreliable for face recognition while pre-clustering can increase the accuracy. The representative feature vector of a cluster can be used as input for the Reidentification feature after the cluster is realized. For details about realizing a new cluster see Cluster Realization.

A typical configuration example for pre-clustering is shown below:

{
   "clustering_config": {
       "method": "SIMPLE_AVERAGE",
       "cluster_realization": {
           "min_num_samples": 5,
           "time_limit_ms": 10000
       },
       "save_internal_state": false,
       "start_from_internal_state": false
   },
   "input_stream_configs": [{
       "stream_id": "camera_1",
       "fv_field_selector": {
           "feature_vector_path": "features"
       },
       "reg_stream_config": {
           "reg_threshold": 0.8,
           "cluster_retention_period_ms": 86400000,
       },
   }]
}

For details about the Reidentification feature see Reidentification. This feature depends on the Facial Feature Vector Extraction feature.