Dataset Viewer
Auto-converted to Parquet Duplicate
clip_id
string
scenario
string
lookalike_of
string
incident
string
category
string
event_frames
list
description
string
scene_domain
string
reference
image
frames
images list
frame_interval_s
float32
label_source
string
generator
string
C0013
camera_cleaning
camera_obstruction
no
none
[]
cleaning a wall camera with a long duster
residential
0.5
Claude visual review; human check pending
ChatGPT image generation
C0029
camera_cleaning
camera_obstruction
no
none
[]
cleaning a wall camera with a long duster
gym
0.5
Claude visual review; human check pending
ChatGPT image generation
C0061
camera_cleaning
camera_obstruction
no
none
[]
cleaning a wall camera with a long duster
loading_dock
0.5
Claude visual review; human check pending
ChatGPT image generation
C0252
camera_obstruction
camera_cleaning
yes
camera
[ 9, 15 ]
man covers the camera with cloth
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0012
camera_obstruction
camera_cleaning
yes
camera
[ 12, 15 ]
man covers the camera with cloth
hotel
0.5
Claude visual review; human check pending
ChatGPT image generation
C0028
camera_obstruction
camera_cleaning
yes
camera
[ 12, 15 ]
man covers the camera with cloth
university
0.5
Claude visual review; human check pending
ChatGPT image generation
C0291
crouching
person_down
no
none
[]
man drops papers and picks them up
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0003
crouching
person_down
no
none
[]
drops papers and crouches to pick them up
industrial
0.5
Claude visual review; human check pending
ChatGPT image generation
C0019
crouching
person_down
no
none
[]
crouches to pick up dropped papers
plaza
0.5
Claude visual review; human check pending
ChatGPT image generation
C0009
greeting
push
no
none
[]
handshake and greeting
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0041
greeting
push
no
none
[]
two people greet and hug
office
0.5
Claude visual review; human check pending
ChatGPT image generation
C0057
greeting
push
no
none
[]
two people greet and hug; last frame jumps to a person close to the camera
warehouse
0.5
Claude visual review; human check pending
ChatGPT image generation
C0111
intact_reflections
null
no
none
[]
reflections in a glass wall
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0031
intact_reflections
null
no
none
[]
reflections in a glass wall
airport
0.5
Claude visual review; human check pending
ChatGPT image generation
C0047
intact_reflections
null
no
none
[]
reflections in a glass wall
library
0.5
Claude visual review; human check pending
ChatGPT image generation
C0026
item_at_jacket
item_into_bag
yes
theft
[ 9, 12 ]
box pushed inside jacket
parking
0.5
Claude visual review; human check pending
ChatGPT image generation
C0042
item_at_jacket
item_into_bag
yes
theft
[ 9, 12 ]
box tucked inside jacket
transit
0.5
Claude visual review; human check pending
ChatGPT image generation
C0058
item_at_jacket
item_into_bag
yes
theft
[ 9, 12 ]
box tucked inside jacket
industrial
0.5
Claude visual review; human check pending
ChatGPT image generation
C0299
item_into_bag
item_at_jacket
no
none
[]
item put into a carried bag
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0011
item_into_bag
item_at_jacket
no
none
[]
box put into carried basket
industrial
0.5
Claude visual review; human check pending
ChatGPT image generation
C0171
item_into_bag
item_at_jacket
no
none
[]
box put into carried wire basket
construction
0.5
Claude visual review; human check pending
ChatGPT image generation
C0103
ordinary_vapor
smoke
no
none
[]
steam from coffee urn
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0007
ordinary_vapor
smoke
no
none
[]
steam from coffee urn
pharmacy
0.5
Claude visual review; human check pending
ChatGPT image generation
C0023
ordinary_vapor
smoke
no
none
[]
uses phone by coffee station
airport
0.5
Claude visual review; human check pending
ChatGPT image generation
C0002
person_down
crouching
yes
violence
[ 9, 15 ]
man collapses to floor, lies there
warehouse
0.5
Claude visual review; human check pending
ChatGPT image generation
C0018
person_down
crouching
yes
violence
[ 6, 15 ]
man sits down then lies on the floor
parking
0.5
Claude visual review; human check pending
ChatGPT image generation
C0034
person_down
crouching
yes
violence
[ 9, 15 ]
man falls and lies on the floor, someone bends over him
transit
0.5
Claude visual review; human check pending
ChatGPT image generation
C0197
phone_in_hand
null
no
none
[]
man walks in and checks his phone
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0021
phone_in_hand
null
no
none
[]
waiting, using phone
gym
0.5
Claude visual review; human check pending
ChatGPT image generation
C0037
phone_in_hand
null
no
none
[]
uses phone in a gallery
museum
0.5
Claude visual review; human check pending
ChatGPT image generation
C0056
push
greeting
yes
violence
[ 9, 12 ]
one man shoves another
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0008
push
greeting
yes
violence
[ 9, 10 ]
one man shoves another in the chest
cinema
0.5
Claude visual review; human check pending
ChatGPT image generation
C0024
push
greeting
yes
violence
[ 9, 12 ]
one man shoves another
sports_concourse
0.5
Claude visual review; human check pending
ChatGPT image generation
C0001
routine
null
no
none
[]
routine counter visit
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0017
routine
null
no
none
[]
customer pays at the till
grocery
0.5
Claude visual review; human check pending
ChatGPT image generation
C0033
routine
null
no
none
[]
routine counter visit
office
0.5
Claude visual review; human check pending
ChatGPT image generation
C0006
smoke
ordinary_vapor
yes
damage
[ 9, 15 ]
smoke rising from bin
loading_dock
0.5
Claude visual review; human check pending
ChatGPT image generation
C0022
smoke
ordinary_vapor
yes
damage
[ 9, 15 ]
smoke rising from bin, people stand up
construction
0.5
Claude visual review; human check pending
ChatGPT image generation
C0038
smoke
ordinary_vapor
yes
damage
[ 9, 15 ]
smoke rising from bin
arcade
0.5
Claude visual review; human check pending
ChatGPT image generation
C0064
uncertain_view
null
no
none
[]
person passes close to the camera
bank
0.5
Claude visual review; human check pending
ChatGPT image generation
C0016
uncertain_view
null
no
none
[]
person walks close past the camera, blurred
cinema
0.5
Claude visual review; human check pending
ChatGPT image generation
C0032
uncertain_view
null
no
none
[]
person walks close past the camera
sports_concourse
0.5
Claude visual review; human check pending
ChatGPT image generation

Lookalikes v1

Short synthetic security-camera sequences in matched pairs. In each pair, one sequence shows an incident and the other shows a harmless scene that looks like it:

Incident Harmless lookalike
Smoke Ordinary vapour or steam
Person down Someone crouching
Push Greeting
Camera covered Camera being cleaned
Item hidden in a jacket Item put into a bag

It also has harmless scenes that often fool camera AI: routine visits, someone holding a phone, glass reflections, and a hard viewing angle.

The set is for evaluation: does a model flag the incident and leave the lookalike alone? It is a small sample published by Knolt.

What's in it

  • 42 sequences across 14 scenarios, 3 each, from 21 kinds of place (bank, pharmacy, transit, warehouse, and others).
  • 15 sequences show an incident; 27 are harmless.
  • Each sequence has:
    • a reference image: the usual view of the scene, nominally 30 s before
    • 16 frames, nominally 0.5 s apart (7.5 s)
  • Frames are small, about 370 × 206 px, like a low-resolution camera stream.

Files:

  • data/test-*.parquet: the set with its images, for the Hub viewer and datasets.
  • clips/<clip_id>/reference.png and clips/<clip_id>/frames/00.png to 15.png: the same images as PNG files.
  • metadata.jsonl: one line per sequence.
  • SHA256SUMS: a checksum for every image.

Fields

Field Meaning
clip_id Sequence ID
scenario What the scene shows, such as smoke or ordinary_vapor
lookalike_of The matching scenario in its pair, if it has one
incident yes or no
category Incident type (violence, fire, theft, camera), or none
event_frames First and last frame where the incident is visible; empty for no
description One line on what happens
scene_domain The kind of place
reference, frames The images
frame_interval_s Nominal time between frames (0.5 s)
label_source, generator How the label and the images were made

Usage

from datasets import load_dataset

ds = load_dataset("KnoltSecurity/lookalike-v1", split="test")
row = ds[0]
row["incident"], row["scenario"], row["frames"][0]

A model is right when it says yes on incident: yes and no on incident: no. Report the two separately: catching incidents and avoiding false alarms are different skills.

How it was made

  • Images: generated with ChatGPT image generation from written prompts. Each sequence came from one generated sheet: a reference panel followed by 16 ordered panels, cut into frames at their native size and not enlarged.
  • People: all fictional adults with small, low-detail faces. No children, no real people, no readable signs or personal data.
  • Labels: from a visual review by Claude. A check of every label by a person at Knolt is in progress; until it's done, treat the labels as provisional.
  • Not included: visible weapons (the image generator refused them) and broken glass (its labels were unclear).
  • These sequences are drawn from the training split of Knolt's Talos 1.1 model; none come from Knolt's held-out benchmark sequences. Don't use this set to claim results for Talos 1.1.

Limits

  • Synthetic. These are still images made to look like a sequence, not recorded video. Motion and timing are nominal, and people or objects can change slightly between frames.
  • Low resolution, by design.
  • Small. 3 sequences per scenario is enough to compare models and spot failure modes, not to estimate accuracy precisely.
  • Not for training a product or making claims about real-world accuracy. Use real recorded footage for that.

Responsible use

  • A camera flag is a candidate for a person to review, never a verdict.
  • This set has no identities and must not be used to identify people, or to infer age, gender or ethnicity.

License

CC BY-NC 4.0: free for research and non-commercial use, with attribution to Knolt.

Citation

@misc{knolt_lookalikes_v1,
  title  = {Lookalikes v1: synthetic security-camera incidents and their harmless lookalikes},
  author = {Knolt},
  year   = {2026},
  url    = {https://huggingface.co/datasets/KnoltSecurity/lookalike-v1}
}
Downloads last month
-