RT-DETR

The RT-DETR filter performs real-time object detection using the RT-DETR (Real-Time DEtection TRansformer) architecture, with IoU-based tracking built in so detections carry a stable identity from frame to frame rather than being re-discovered on every frame.

Features

• Real-time detection — transformer-based multi-class bounding box extraction.
• Built-in IoU tracking — associates detections across frames by bounding-box overlap, so downstream filters can act per object instead of per frame.
• Batched inference — processes frames in batches, with sparse-mode batching gated on whether object tracking is enabled.
• Configurable thresholds — confidence and overlap thresholds tuned per pipeline.

Use cases

• Detecting and following objects through a live video stream.
• Feeding stable object identities to attribute or re-identification filters downstream.
• High-throughput detection where per-frame re-detection would be wasteful.