What it is🔗
A ROS2 package that wraps a TensorRT engine (YOLOv11 detect and seg) and runs inference on a static image in a loop, saving:
- bounding boxes (and optionally masks)
- annotated output images
- logs/parameters used for the run
Highlights🔗
- Correct post-processing: logits vs sigmoid, xywh decoding, scaling, thresholds
- Stable filtering:
pre_topk/max_det, confidence cutoffs, deterministic output - Config-driven runs: launch parameters control inputs/outputs without recompiling
- Supports multiple model variants: detection-only and segmentation engines
Why it matters🔗
This is meant for deployment-style workflows: predictable results, easy parameter tuning, and repeatable runs that can be used in robotics pipelines (ROS2 nodes, launch files, and CI test images).