Auspex RT
betav0.1.0Bounding-box object detection, trained from scratch on your own data — one task, with nothing else taking up parameters or compute.
Install
pip install auspex-rtOverview
Auspex RT trains a single object detector on your bounding boxes and runs it in your own stack. Nothing runs on this site.
It is built from random initialisation. No pretrained weights are downloaded and none are required, so the detector you train carries no upstream weights and no licence attached to them.
It reads YOLO and Roboflow text exports, COCO JSON, CVAT-for-images 1.1 XML and Pascal VOC, mixed freely in one training run. A dataset annotated with polygons still trains a box model: the box is taken from the annotation's extent, and the conversion is counted and reported rather than done silently.
Everything that is not detection was left out rather than switched off — no segmentation head, no keypoint head, no prototype masks. That makes it a smaller network at roughly half the compute of a general-purpose multi-task model at the same input size, and it means there is no configuration in which you are paying for a head you cannot train.
Predictions export back to CVAT XML or COCO JSON for pre-labelling, very large images can be tiled so small objects stay detectable, and per-class confidence thresholds are calibrated into the checkpoint so classes with different score scales each show up at their own best operating point.
- Bounding boxes only — no unused heads consuming parameters or compute
- Trained from random init: no pretrained weights, nothing downloaded
- Reads YOLO/Roboflow, COCO, CVAT and Pascal VOC exports in one run
- Polygon annotations train a box model; the conversion is reported, never silent
- Per-class confidence calibration and tiled inference for large images
- Licence verified offline — training works on air-gapped machines
Need this adapted to your data?
We build and evaluate models against your metrics, not benchmarks.