Visual classification models classify an entire image into different categories by assigning a confidence score for each class.
Classification models can be multi-headed, where each group of mutually exclusive model classes belong to a single model head. For example, when an image is run through Hive's visual moderation model, one head might classify sexually not-safe-for-work (NSFW) content while another head might classify the presence of guns.
This concept is illustrated below. This imaginary model has two heads:
NSFW classification: general_nsfw, general_suggestive, general_not_nsfw_not_suggestive
Gun classification: gun_in_hand, animated_gun, gun_not_in_hand, no_gun
The confidence scores for each model head sum to 1.
When submitting a video to be processed, Hive’s backend splits the video into frames, runs the model on each frame, then recombines the results into an aggregated response for the entire video. The video output for a classifier is similar to a list of classification output objects, but with multiple timestamps.
A more detailed walkthrough on how to submit visual classification tasks via the API and how to interpret the visual model response can be found in our customer guide.
Hive's visual classification models support a wide variety of classes that are relevant to content moderation. Broadly, visual moderation classes can be separated into five main categories: sexual content, violent imagery, drugs, hate imagery, and image attributes. When deciding how to process our API response in order to implement your content policy, you should consult the following class descriptions to decide which classes to moderate.
Note: Older versions of the API might not perfectly match the outline below. Please reach out to [email protected] if you would like to access the latest content moderation classes.
- general_nsfw - genitalia, sexual activity, nudity, buttocks, sex toys, animal genitalia
- general_suggestive - shirtless men, underwear / swimwear, sexually suggestive poses without genitalia, occluded or blurred sexual activity
- general_not_nsfw_not_suggestive - none of the above, clean
Sexual Activity Head:
- yes_sexual_activity - a sex act or stimulation of genitals are present in the scene
- no_sexual_activity - no sex act is present in the scene
Realistic NSFW Head:
- yes_realistic_nsfw - live nudity, sex acts, or photo-realistic representations of nudity or sex acts
- no_realistic_nsfw - non-photorealistic representations of nudity or sex acts (statues, crude drawings, paintings etc.); lack of any NSFW content
Female Underwear Head:
- yes_female_underwear - lingerie, bras, panties
Male Underwear Head:
- yes_male_underwear - fruit-of-the-loom, boxers
Sex Toy Head:
- yes_sex_toy - dildos, certain lingerie
Female Nudity Head:
- yes_female_nudity - breasts or female genitalia
Male Nudity Head:
- yes_male_nudity - male genitalia
Female Swimwear Head:
- yes_female_swimwear - bikinis, one-pieces, not underwear
Shirtless Male Head:
- yes_male_shirtless - shirtless below mid-chest
Sexual Intent Head: (beta)
- yes_sexual_intent - occluded, blurred, or hidden sexual activity
Animal Genitalia Head: (beta)
- animal_genitalia_and_human - sexual activity including both animals and humans
- animal_genitalia_only - animals mating and pictures of animal genitalia
- animated_animal_genitalia - drawings of sexual activity involving animals
- no_animal_genitalia - none of the above, clean
- gun_in_hand - person holding rifle, handgun
- gun_not_in_hand - rifle, handgun, not in hand
- animated_gun - gun in games, cartoons, etc. can be in-hand or not.
- knife_in_hand - person holding knife, sword, machete, razor blade
- knife_not_in_hand - knife, sword, machete, razor blade, not in hand
- culinary_knife_in_hand - knife being used for preparing food
- very_bloody - gore, visible bleeding, self-cutting
- a_little_bloody - fresh cuts / scrapes, light bleeding
- no_blood - minor scabs, scars, acne, etc. are not considered ‘blood’ by model
- other_blood - animated blood, fake blood, animal blood such as game dressing
- hanging - the presence of a human hanging by noose (dead or alive)
- noose - a noose is present in the image with no human hanging from it
- no_hanging_no_noose - no person hanging and no noose present
Corpses Head: (beta)
- human_corpse: human dead body present in image
- animated_corpse: animated dead body present in image
Emaciated Bodies Head:
- yes_emaciated_body: emaciated human or animal body present in image
Self Harm Head: (beta)
- yes_self_harm: self cutting, burning, instances of suicide or other self harm methods present in image
- yes_pills - pills and / or drug powders
- no_pills - no pills and / or drug powders
- illicit_injectables - heroin and other illegal injectables
- medical_injectables - injectables for medical use
- no_injectables - no injectable drug paraphernalia
- yes_smoking - cigarettes, cigars, marijuana, vapes, or other smoking paraphernalia
- no_smoking - no cigars, marijuana, vapes, or other smoking paraphernalia
- yes_nazi - Nazi symbols
- no_nazi - absence of the above
- yes_terrorist - ISIS flag
- no_terrorist - absence of the above
White Supremacy Head:
- yes_kkk - KKK symbols
- no_kkk - absence of the above
Middle Finger Head:
- yes_middle_finger - middle finger
- no_middle_finger - absence of the above
- text - any form of text or writing is present somewhere on the image
- no_text - no text present in the image
Overlay Text Head:
- yes_overlay_text - digitally overlaid text is present on an image (think meme text)
- no_overlay_text - lack of digitally overlaid text in the image
- yes_child_present: a baby or toddler is present in the image
- yes_drawing: a drawing, painting, or sketch is the central part of the image
Image Type Head:
- animated - the image is animated
- hybrid - the image is partially animated
- natural - the image has no animation
If you need more information when deciding which classes to use, a comprehensive list of subject matter covered by each visual class is available in this visual taxonomy document (warning: somewhat NSFW).
Hive's Brand Safety and Brand Suitability APIs are powered by Hive's visual moderation model and are additionally mapped to the GARM Brand Safety & Suitability Framework (Global Alliance for Responsible Media), which was established as an industry-standard for categorizing harmful content. For more information click here for more information.
For each of the classes mentioned above, you will need to set thresholds to decide when to take action based on our model results. For optimum results, a proper threshold analysis on a natural distribution of your data is recommended (for more on this please contact Hive at the email below). Generally, though, a model confidence score threshold of >.90 is a good place to start to flag an image for any class of interest.
For questions on best practices, please message your point of contact at Hive or send a message to [email protected] to contact our API team directly.
Updated 16 days ago