Independent ECDet-S FP32 ONNX conversion for AnnotateIt
Hi EdgeCrafter team,
I’ve published an independent, unofficial FP32 ONNX conversion of ECDet-S for local inference in AnnotateIt:
https://huggingface.co/AnnotateIt/edgecrafter-ecdet-s-onnx
The repository includes the pinned upstream source, Apache-2.0 attribution,
export and verification scripts, the preprocessing/output contract, SHA-256
checksums, and PyTorch ↔ ONNX Runtime validation results.
The model preserves the raw EdgeCrafter outputs:
pred_logits [batch, 300, 80] and pred_boxes [batch, 300, 4].
Integration into AnnotateIt is currently undergoing final validation.
Upstream feedback on the conversion is very welcome.
@Xiiiiii0220 , Hi, I’ve just published a write-up about using EdgeCrafter for local pre-labeling in AnnotateIt:
https://annotateit.ai/local-computer-vision-pre-labeling/
I tested the ONNX versions of ECDet-S/M and ECSeg-S/M in a real browser-based annotation workflow, including a couple of unsuccessful INT8 and WebGPU experiments.
If you have a few minutes, I’d really appreciate a quick sanity check, especially on how I’ve described the original models and results. And if you think the article would be useful to the EdgeCrafter community, I’d be happy to adapt it for Hugging Face with you.