Thanks for CommonForms, the dataset and the models are a genuinely useful release.
The model cards for FFDNet-S and FFDNet-L do not state a license, and there is no LICENSE file in this repo, so as things stand there is no grant that would let anyone redistribute the weights. FFDetr is tagged apache-2.0 on the Hub, which is what we use today (exported to Core ML, running on-device in a Mac app to auto-place a signature on a form). It does well on text and choice fields, but its Signature scores are much weaker than the 93.5 AP you report for FFDNet-L, which is the class we care about most.
Two questions:
- Is an explicit license planned for the FFDNet-S / FFDNet-L weights? Apache-2.0 or MIT would let us use them directly.
- Since the FFDNet detectors are YOLO11 and Ultralytics is AGPL-3.0, is that what kept them unlicensed while FFDetr went out under Apache-2.0? If the answer is simply "not possible", that is a useful answer too, and no hard feelings.
Either way, thank you for putting the dataset out under a permissive license. If the weights cannot be licensed we will train our own detector on CommonForms instead.
Thanks for CommonForms, the dataset and the models are a genuinely useful release.
The model cards for FFDNet-S and FFDNet-L do not state a license, and there is no LICENSE file in this repo, so as things stand there is no grant that would let anyone redistribute the weights. FFDetr is tagged
apache-2.0on the Hub, which is what we use today (exported to Core ML, running on-device in a Mac app to auto-place a signature on a form). It does well on text and choice fields, but its Signature scores are much weaker than the 93.5 AP you report for FFDNet-L, which is the class we care about most.Two questions:
Either way, thank you for putting the dataset out under a permissive license. If the weights cannot be licensed we will train our own detector on CommonForms instead.