Vegetables as porn?

[Saymon1998]

Hi everyone, lately I came into an issue where mobilenetV1 model (2,6 MB) gave me most strange output when I send an image with onions to classify. I did check it also on other vegetables, like apples and beetroot, and it looks like that only apples aren't classified as porn/hentai :) Onions gave me results: 0: {className: "Porn", probability: 0.7866919040679932} 1: {className: "Neutral", probability: 0.14794990420341492} 2: {className: "Sexy", probability: 0.04114266857504845} 3: {className: "Hentai", probability: 0.022164924070239067} 4: {className: "Drawing", probability: 0.0020506654400378466}

Apples: 0: {className: "Neutral", probability: 0.9997453093528748} 1: {className: "Drawing", probability: 0.00016232636698987335} 2: {className: "Porn", probability: 0.00005498077371157706} 3: {className: "Sexy", probability: 0.00002530598248995375} 4: {className: "Hentai", probability: 0.000012140934813942295}

Beetroot: 0: {className: "Hentai", probability: 0.41282495856285095} 1: {className: "Drawing", probability: 0.302909255027771} 2: {className: "Neutral", probability: 0.20993521809577942} 3: {className: "Porn", probability: 0.05640846863389015} 4: {className: "Sexy", probability: 0.01792212948203087}

I wonder, why? Does mobilenet model react like that because of onions colour? Maybe beetroots roots somehow resemble tentacles in some way, at least, for this model?

Used images:

Cheers

[952xuan]

White and reflective Mantous(Chinese steamed buns) can be more pornographic. Maybe we should do human target detection before using this model.

[GantMan]

thank you for this - will add to training