yoloe-26s-seg-onnx

YOLOE-26s-seg — fixed-vocabulary ONNX exports (torch-free runtime)

Fixed-vocabulary ONNX exports of yoloe-26s-seg for torch-free edge inference (onnxruntime CPU / CUDA / TensorRT EP). The class vocabulary and NMS are baked into the graph at export time, so the runtime needs no CLIP/mobileclip text encoder and no ultralytics/torch — just onnxruntime + a small numpy/cv2 seg post-process.

Used by the reBot B601-DM voice-arm vision-grasp pipeline (openvoicestream, app ovs_agent/apps/voice_rebot_arm, see perception/yolo_onnx.py for the matching post-processing and RUNBOOK.md for deployment).

Files

Both: input images [1,3,640,640]; outputs output0 [1,300,38] (= 4 xyxy + conf + cls_id + 32 mask coeffs, NMS already applied in-graph) and output1 [1,32,160,160] (mask prototypes).

Reproduce the export

from ultralytics import YOLOE  # needs torch at EXPORT time only
m = YOLOE("yoloe-26s-seg.pt")
m.set_classes([...])  # the vocabulary above — order defines class ids
m.export(format="onnx", opset=12, simplify=True, imgsz=640)

Post-processing gotcha

Ultralytics' masks.data strips the letterbox padding before storing; a numpy re-implementation must strip padding first, then resize to the original frame or mask areas come out systematically ~25% low. See perception/yolo_onnx.py.

TensorRT note (Jetson)

Building a TRT engine from these (ORT TensorrtExecutionProvider) peaks high in memory during tactic search — pre-build once in a free-memory window with trt_engine_cache_enable, then the cached engine loads fine under memory pressure.