I’m evaluating semantic segmentation models like SegNeXt and PP‑LiteSeg, which are released under the Apache 2.0 license. The code is clearly permissive, but I’m unsure about the pretrained weights.
Many segmentation models are trained on datasets such as Cityscapes, which has non‑commercial restrictions. My concern is:
When a model repo says it’s Apache 2.0 licensed, does that license also cover the released pretrained weights?
If the weights were trained on a non‑commercial dataset (like Cityscapes), does that mean I cannot use those weights in a commercial product, even if the code is Apache 2.0?
Are SegNeXt and PP‑LiteSeg safe to use commercially with their provided weights, or do I need to retrain them on a fully commercial‑friendly dataset?
I’d like clarification from people who have dealt with licensing in ML projects. How do you interpret the license vs dataset restrictions in practice?