Detection of the larger model is worse

[aios-ai]

I expected the larger model, 640m.onnx, to detect more compared to the 320n.onnx model, albeit with a longer processing time. While it indeed takes longer, I was surprised to find that it detects fewer nudes. In my test with around 1,000 images, the large model detected 640 instances, whereas the smaller model detected 707. The large model clearly missed some images that were obviously nude. It is not due to improved false positives. Has anyone else experienced this issue or have any tips on how to investigate it?

[bedapudi6788]

@Sascha353 hi, still in the process of testing the models and also adding some more models. I will be able to give full clarity after few days

[mpisat]

How is it compared to original dataset? I mean detector_v2_default