📕
Items and Characteristics
Automatically check any videos and images for harmful content
The ITEMS_CHARACTERISTICS endpoints surface items and characteristics detected in an image (object recognition) or video (including sound). Each item or characteristic has a prevalence score which represents the extent to which it is present in a given piece of content. We can also analyse the text that is shared along the image or video.
You can use these scores in mappings or models to implement safety policies, for example setting a numerical threshold for automated flagging of content or in a linear regression model (read more in our guide for selecting thresholds).
The following 50+ items and characteristics are available:
Violence:
violence
firearm
knife
- knives in all contexts, e.g. includes kitchen knivesviolent_knife
- knives in violent contexts, including hunting knives
Substances:
alcohol
- alcoholic drinks in bottles, cans or glassesdrink
- drinks of any kindssmoking_and_tobacco
pills
recreational_pills
Hate symbols:
confederate_flag
pepe_frog
nazi_swastika
NSFW:
adult_content
suggestive
adult_toys
Visual content characteristics:
artistic
comic
meme
photo
screenshot
map
poster_cover
game_screenshot
face_filter
promo_info_graphic
Audio language toxicity:
toxic
severe_toxic
obscene
insult
identity_hate
threat
sexual_explicit
OCR language toxicity:
toxic
severe_toxic
obscene
insult
identity_hate
threat
sexual_explicit
Caption language toxicity:
toxic
severe_toxic
obscene
insult
identity_hate
threat
sexual_explicit
Other:
middle_finger_gesture
child
toy
gambling_machine
The classes below have been developed to beta standard. They can be included in offline evaluations and will be available as add-on features in early October 2023
🙋🏽♀️ Note: content containing these items and characteristics can still be detected with Custom policies and with our Brand Safety Framework. This is because these products use inputs that represent all aspects of the content - not just Items & Characteristics. For example, our GARM product flags harmful White Supremacy content as Hate Speech through learnt patterns in the content, even though this isn’t an Item/Characteristic.
NSFW:
medical
- NSFW content in a medical context, e.g. partial nudity in a breast examinationover_18
exposed_anus
exposed_armpits
exposed_belly
covered_belly
covered_buttocks
exposed_buttocks
face_f
face_m
covered_feet
exposed_feet
covered_breast_f
exposed_breast_f
covered_genitalia_f
exposed_genitalia_f
exposed_breast_m
exposed_genitalia_m
In early October 2023, we'll be including two add-on features in our Items & Characteristics response. Please let us know at [email protected] if you'd be interested in having these:
- 1.Optical Character Recognition (OCR). OCR refers to any text that appears on the image or video. Examples include the captions for translations, words displayed on a T-shirt or hhandwritten content.
- 2.Speech audio transcriptions. A literal transcription of the speech detected during a video. This is only available in English. More languages are coming up soon, starting with Spanish. Please be aware that any sound in the video may interfere with the audio transcription.
You can find below an example of how to process OCR and audio transcriptions from the Items & Characteristics API response. They are included in the sections "ocr_texts" and "audio_texts".
{
"XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX": {
"status": "done",
"results": {
"violence": {
"violence": 0.0263,
"firearm": 0.0673,
"knife": 0.102,
"violent_knife": 0.0774
},
[...]
"other": {
"child": 0.5896,
"middle_finger_gesture": 0.0291,
"toy": 0.0278,
"gambling_machine": 0.0595
},
"url": "https://whatever.com/test-video_itemschar.mp4",
"ocr_texts": [
"I'm very happy",
],
"metadata": {
"width": 576,
"height": 1024,
"fps": 30.0,
"duration": 10.134,
"seconds_processed": 10.1
},
"audio_texts": [
"I've been thinking about it and I'm happy"
]
},
"msg": null
}
}
Last modified 3h ago