DEIMv2 Detection

The DEIMv2 Detection filter performs real-time object detection using the DEIMv2 (DETR with Improved Matching, v2) architecture. It processes frames from one or more topics, extracts multi-class bounding boxes, groups them as regions of interest, and records per-frame confidence metrics for downstream filters.

Features

• Multi-topic detection — runs detection independently across several input topics.
• Configurable NMS — standard, weighted, or soft non-maximum suppression to resolve overlapping detections.
• Result aggregation — combines detections across frames, with per-class ignore lists for classes a pipeline does not care about.
• Optional overlay — renders bounding boxes and class labels onto the frame when visualization is enabled.

Use cases

• Enriching streaming video with object detections for downstream analytics.
• Extracting regions of interest for a cropping or classification filter further along the pipeline.
• Tagging frames with detection metadata for search and review.