Dataset Preview · KISA2016
ETRI CCTV Anomaly Event Dataset
Fixed-camera surveillance videos for anomaly detection. Every video has one XML annotation giving the event class, onset time, and duration.
6event categories
100videos
9.6 htotal footage
1280×720resolution @ 30 fps
37night-time videos
| Category | Event | Videos | Footage | Night |
|---|---|---|---|---|
| Abandonment | An object is left behind and its owner walks away | 10 | 00:52:20 | 3 |
| Arson | A person sets a fire in the scene | 10 | 01:01:01 | 4 |
| Falldown | A person collapses and stays on the ground | 10 | 00:54:45 | 3 |
| Intrusion | A person enters a restricted area | 30 | 02:45:07 | 8 |
| Loitering | A person lingers in the area for an extended time | 30 | 03:04:00 | 14 |
| Violence | Physical fighting between people | 10 | 01:00:23 | 5 |
Abandonment
An object is left behind and its owner walks away · 10 videos · 00:04:00 – 00:06:37 each
C00_015_0001.mp4
ScenarioAbandonment
Dataset / StageKISA2016 / 2
LocationTraditionalHouse
Time of dayDay
WeatherClear
EventAbandonment · onset 00:03:36 · lasts 10 s
00:00:00
alarm 00:03:36 + 10 s
00:06:36
C00_023_0001.mp4
ScenarioAbandonment
Dataset / StageKISA2016 / 2
LocationInsideBuiliding
Time of dayDay
WeatherClear
EventAbandonment · onset 00:03:34 · lasts 10 s
00:00:00
alarm 00:03:34 + 10 s
00:06:37
C00_028_0001.mp4
ScenarioAbandonment
Dataset / StageKISA2016 / 2
LocationAlley
Time of dayDay
WeatherClear
EventAbandonment · onset 00:03:31 · lasts 10 s
00:00:00
alarm 00:03:31 + 10 s
00:04:32
C00_036_0001.mp4
ScenarioAbandonment
Dataset / StageKISA2016 / 3
LocationPlayground
Time of dayDawn
WeatherClear
EventAbandonment · onset 00:03:29 · lasts 10 s
00:00:00
alarm 00:03:29 + 10 s
00:06:36
Arson
A person sets a fire in the scene · 10 videos · 00:03:30 – 00:08:39 each
C00_012_0007.mp4
ScenarioFireDetection
Dataset / StageKISA2016 / 1
LocationTraditionalHouse
Time of dayDay
WeatherClear
EventFireDetection · onset 00:03:44 · lasts 20 s
00:00:00
alarm 00:03:44 + 20 s
00:08:39
C00_038_0006.mp4
ScenarioFireDetection
Dataset / StageKISA2016 / 1
LocationRoad
Time of dayDusk
WeatherClear
EventFireDetection · onset 00:03:28 · lasts 20 s
00:00:00
alarm 00:03:28 + 20 s
00:07:23
C00_049_0001.mp4
ScenarioFireDetection
Dataset / StageKISA2016 / 1
LocationFactory
Time of dayDay
WeatherClear
EventFireDetection · onset 00:03:27 · lasts 20 s
00:00:00
alarm 00:03:27 + 20 s
00:07:32
C00_049_0002.mp4
ScenarioFireDetection
Dataset / StageKISA2016 / 1
LocationFactory
Time of dayDay
WeatherClear
EventFireDetection · onset 00:03:45 · lasts 20 s
00:00:00
alarm 00:03:45 + 20 s
00:07:44
Falldown
A person collapses and stays on the ground · 10 videos · 00:03:44 – 00:07:57 each
C00_003_0003.mp4
ScenarioFalldown
Dataset / StageKISA2016 / 1
LocationLocalFacilities
Time of dayDay
WeatherClear
EventFalldown · onset 00:03:18 · lasts 10 s
00:00:00
alarm 00:03:18 + 10 s
00:07:53
C00_023_0004.mp4
ScenarioFalldown
Dataset / StageKISA2014 / 1
LocationInsidebuiliding
Time of dayDay
WeatherClear
EventFalldown · onset 00:03:17 · lasts 10 s
00:00:00
alarm 00:03:17 + 10 s
00:07:37
C00_034_0005.mp4
ScenarioFalldown
Dataset / StageKISA2016 / 2
LocationPlayground
Time of dayDusk
WeatherClear
EventFalldown · onset 00:03:08 · lasts 10 s
00:00:00
alarm 00:03:08 + 10 s
00:07:19
C00_060_0003.mp4
ScenarioFalldown
Dataset / StageKISA2016 / 1
LocationBusstation
Time of dayDay
WeatherClear
EventFalldown · onset 00:03:31 · lasts 10 s
00:00:00
alarm 00:03:31 + 10 s
00:04:30
Intrusion
A person enters a restricted area · 30 videos · 00:03:00 – 00:10:45 each
C00_005_0001.mp4
ScenarioIntrusion
Dataset / StageKISA2016 / 2
LocationLocalFacilities
Time of dayDawn
WeatherClear
EventIntrusion · onset 00:04:15 · lasts 8 s
00:00:00
alarm 00:04:15 + 8 s
00:07:23
C00_008_0002.mp4
ScenarioIntrusion
Dataset / StageKISA2016 / 3
LocationLocalFacilities
Time of dayDay
WeatherClear
EventIntrusion · onset 00:04:07 · lasts 10 s
00:00:00
alarm 00:04:07 + 10 s
00:07:17
C00_014_0001.mp4
ScenarioIntrusion
Dataset / StageKISA2016 / 2
LocationTraditionalHouse
Time of dayDawn
WeatherClear
EventIntrusion · onset 00:03:23 · lasts 4 s
00:00:00
alarm 00:03:23 + 4 s
00:05:24
C00_014_0002.mp4
ScenarioIntrusion
Dataset / StageKISA2016 / 2
LocationTraditionalHouse
Time of dayDay
WeatherClear
EventIntrusion · onset 00:03:17 · lasts 2 s
00:00:00
alarm 00:03:17 + 2 s
00:06:51
Loitering
A person lingers in the area for an extended time · 30 videos · 00:03:50 – 00:08:20 each
C00_001_0001.mp4
ScenarioLoitering
Dataset / StageKISA2016 / 1
LocationLocalFacilities
Time of dayDay
WeatherClear
EventLoitering · onset 00:03:00 · lasts 64 s
00:00:00
alarm 00:03:00 + 64 s
00:07:04
C00_004_0001.mp4
ScenarioLoitering
Dataset / StageKISA2016 / 2
LocationLocalFacilities
Time of dayDay
WeatherClear
EventLoitering · onset 00:03:18 · lasts 89 s
00:00:00
alarm 00:03:18 + 89 s
00:07:50
C00_004_0004.mp4
ScenarioLoitering
Dataset / StageKISA2016 / 2
LocationLocalFacilities
Time of dayDawn
WeatherClear
EventLoitering · onset 00:03:18 · lasts 75 s
00:00:00
alarm 00:03:18 + 75 s
00:07:39
C00_010_0001.mp4
ScenarioLoitering
Dataset / StageKISA2016 / 1
LocationTraditionalHouse
Time of dayDawn
WeatherClear
EventLoitering · onset 00:03:18 · lasts 63 s
00:00:00
alarm 00:03:18 + 63 s
00:07:26
Violence
Physical fighting between people · 10 videos · 00:03:50 – 00:07:40 each
C00_015_0003.mp4
ScenarioViolence
Dataset / StageKISA2016 / 2
LocationTraditionalHouse
Time of dayDawn
WeatherClear
EventViolence · onset 00:03:28 · lasts 20 s
00:00:00
alarm 00:03:28 + 20 s
00:07:35
C00_020_0001.mp4
ScenarioViolence
Dataset / StageKISA2016 / 2
LocationSchoolbuilding_1
Time of dayDawn
WeatherClear
EventViolence · onset 00:03:16 · lasts 20 s
00:00:00
alarm 00:03:16 + 20 s
00:07:11
C00_032_0002.mp4
ScenarioViolence
Dataset / StageKISA2016 / 1
LocationPlayground
Time of dayDay
WeatherClear
EventViolence · onset 00:03:11 · lasts 20 s
00:00:00
alarm 00:03:11 + 20 s
00:06:55
C00_038_0003.mp4
ScenarioViolence
Dataset / StageKISA2016 / 1
LocationRoad
Time of dayDay
WeatherClear
EventViolence · onset 00:03:17 · lasts 20 s
00:00:00
alarm 00:03:17 + 20 s
00:07:06
Annotations: the XML files
Modalities: video + XML only. The dataset contains exactly 100 mp4 files and
100 same-named XML annotations — there is no QA, caption, transcript, or any other text. Every sample
card above renders its own XML in full; the table below documents what each field means.
Field reference
| Field | Meaning | Example values |
|---|---|---|
| Scenario | Event category of the clip | Abandonment, Arson, Falldown, Intrusion, Loitering, Violence |
| Filename / Duration | Video file name and full length | C00_015_0001.mp4 · 00:06:36 |
| AlarmEvents | Number of annotated events in the clip | 1 for every clip in this set |
| Location | Shooting site — inconsistently filled (see note below) | Alley, Basstation, Beachside, Busstation, Factory, InsideBuiliding, … / C080, C081, C082, C086, … |
| Weather | TimeOfDay (Day/Night), Clouds, Windy, Rain, Snow, Fog | Day, None, None, No, No, No |
| DetectArea | Polygon of the region monitored for the event | 0,0 → 1280,0 → 1280,720 → 0,720 (full frame) |
| Alarm / StartTime | When the event begins | 00:03:36 |
| Alarm / AlarmDescription | Event class of this alarm (matches Scenario) | Abandonment |
| Alarm / AlarmDuration | How long the event lasts | 00:00:10 |
Caveat — the Location field is inconsistent. 50 of the 100 clips use semantic
scene names (Alley, Basstation, Beachside, Busstation, Factory, InsideBuiliding, Insidebuiliding, Insidefactory, LocalFacilities, OutdoorParking, Park, Parking_1, Parking_2, PlayGround, Playground, Road, Schoolbuilding, Schoolbuilding_1, Schoolbuilding_2, Stream, TraditionalHouse, Trainplatform — note the typos, e.g. "InsideBuiliding", "Basstation"), while the
other 50 use raw camera-site codes (C080, C081, C082, C086, C087, C090, C092, C095, C096, C097, C098, C099, C100, C102, C103, C104, C105, C106, C107, C110, C111, C113, C114) that do not match the camera number in the
filename and have no lookup table in the release. For scene-based splits, prefer grouping by the camera
number embedded in filenames (e.g. all C00_146_* clips share one camera) over trusting this field.
Pilot evaluation — 12 clips (2 per category), not the full dataset.
Two protocols: streaming models receive frames one per second and must decide each second
whether to stay silent or report the event as it happens; offline models see 64 timestamped
frames of the whole video at once and answer a single question — an after-the-fact upper bound.
Streaming protocol (online detection @ 1 fps)
| Model | Clips | Recall | Recall (±30 s) | Latency | Class acc | False alarms/h |
|---|---|---|---|---|---|---|
| JoyAI-VL-8B | 12 | 17% | 58% | -1.6 s | 43% | 4.7 |
Strict recall counts a inside [onset, end+30 s]. The relaxed column also accepts responses up to 30 s before the annotated onset: the KISA onset marks when the alarm criterion is met (e.g. object unattended for N seconds), and models often report the precursor action earlier. Latency is relative to annotated onset (negative = earlier). Streamo-3B pending (checkpoint unavailable, see notes).
JoyAI-VL-8B — when it speaks vs. the annotated alarm
Orange band = annotated alarm interval; dots = model responses.
Abandonment/C00_015_0001
00:06:36
Loitering/C00_001_0001
00:07:04
Abandonment/C00_023_0001
00:06:37
Loitering/C00_004_0001
00:07:50
Arson/C00_012_0007
00:08:39
Violence/C00_015_0003
00:07:35
Arson/C00_038_0006
00:07:23
Violence/C00_020_0001
00:07:11
Falldown/C00_003_0003
00:07:53
Falldown/C00_023_0004
00:07:37
Intrusion/C00_005_0001
00:07:23
Intrusion/C00_008_0002
00:07:17
Offline protocol (whole-video QA, 64 frames)
| Model | Clips | Parse rate | Class acc | Onset error | tIoU |
|---|---|---|---|---|---|
| Cosmos-Reason1-7B | 12 | 100% | 58% | 12.6 s | 0.13 |
| Qwen2.5-VL-3B | 12 | 100% | 33% | 53.0 s | 0.09 |
| Qwen2.5-VL-7B | 12 | 100% | 92% | 24.4 s | 0.25 |
Extra task probes (Qwen2.5-VL-7B, offline)
Region-conditioned classification
DetectArea polygon given in the prompt; classify the event inside it
class acc 75% (n=12); on non-full-frame regions 86% (n=7)
Hallucination probe (wrong event)
“Does this video contain <event that never happens>?” — ground truth is always No
false-yes rate 0% (n=12, lower is better)
Event verification (true event)
Same question for the event that does occur — ground truth is always Yes
yes rate 42% (n=12, higher is better)
Attribute QA (time of day)
“What time of day was this recorded?” vs the XML Weather/TimeOfDay field
accuracy 92% (n=12)
Case studies — all 12 pilot clips
Each card: the event clip, a unified timeline (orange band = annotated alarm; dots = JoyAI responses; colored bars = offline models' predicted interval), every model's actual answer, and the per-clip task probes. Timelines span the full video, so an offline bar far from the orange band is a temporal-grounding miss even when the class is right.
Abandonment/C00_015_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:06:36
JoyAI-VL-8B
Alert: a person has entered the courtyard area.
Qwen2.5-VL-7B wrong class
A person in a blue jacket walks across the courtyard.
Cosmos-Reason1-7B correct
A person in a blue jacket and black pants is seen walking away from the camera, leaving a green bag behind on the ground.
Qwen2.5-VL-3B wrong class
a person enters a restricted area
Qwen2.5-VL-7B probes — asked about Arson (never happens) → said no · asked about Abandonment (does happen) → said no · time of day: Day (gt Day) ✓
Abandonment/C00_023_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:06:37
JoyAI-VL-8B
A person has fallen down in the hallway.
Qwen2.5-VL-7B correct
A person drops a bag on the floor and walks away.
Cosmos-Reason1-7B correct
The person drops the plastic bag on the floor.
Qwen2.5-VL-3B correct
a person leaves a bag on the floor
Qwen2.5-VL-7B probes — asked about Arson (never happens) → said no · asked about Abandonment (does happen) → said yes · time of day: Day (gt Day) ✓
Arson/C00_012_0007
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:08:39
JoyAI-VL-8B
A person has entered the monitored area from the bottom left of the frame.
JoyAI-VL-8B
A fire has started at the metal container on the wall, with visible smoke rising.
Qwen2.5-VL-7B correct
A person lights a fire on a grill.
Cosmos-Reason1-7B correct
A person approaches a barbecue grill, lights it, and steps back, causing the grill to catch fire.
Qwen2.5-VL-3B wrong class
a man leaves the grill
Qwen2.5-VL-7B probes — asked about Falldown (never happens) → said no · asked about Arson (does happen) → said yes · time of day: Day (gt Day) ✓
Arson/C00_038_0006
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:23
JoyAI-VL-8B
An abandoned object was detected on the road. A black stool was left unattended in the middle of the road.
JoyAI-VL-8B
Arson was detected. A person set fire to a stool on the road, and flames are visible.
Qwen2.5-VL-7B correct
A person sets a fire to a grill on the street.
Cosmos-Reason1-7B correct
A person approaches a barbecue stand and lights it on fire using a white cloth.
Qwen2.5-VL-3B wrong class
a man walks away from the grill
Qwen2.5-VL-7B probes — asked about Falldown (never happens) → said no · asked about Arson (does happen) → said yes · time of day: Dusk (gt Dusk) ✓
Falldown/C00_003_0003
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:53
JoyAI-VL-8B
A person has fallen down near the fence.
Qwen2.5-VL-7B correct
A person falls to the ground and remains motionless.
Cosmos-Reason1-7B wrong class
A person in a red jacket and black pants collapses onto the ground, lying motionless.
Qwen2.5-VL-3B wrong class
a person leaves a red coat on the ground
Qwen2.5-VL-7B probes — asked about Violence (never happens) → said no · asked about Falldown (does happen) → said yes · time of day: Dawn (gt Day) ✗
Falldown/C00_023_0004
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:37
JoyAI-VL-8B
A person has fallen down in the hallway.
Qwen2.5-VL-7B correct
A person falls to the ground in the hallway.
Cosmos-Reason1-7B correct
A person wearing a red jacket and black pants falls backward onto the floor, lying flat with their back to the ground.
Qwen2.5-VL-3B correct
a person falls down
Qwen2.5-VL-7B probes — asked about Violence (never happens) → said no · asked about Falldown (does happen) → said yes · time of day: Day (gt Day) ✓
Intrusion/C00_005_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:23
JoyAI-VL-8B
A person fell down on the sidewalk.
Qwen2.5-VL-7B correct
A person in a red jacket walks along the sidewalk and then bends down to pick something up from the ground.
Cosmos-Reason1-7B wrong class
A person in a red jacket and dark pants is seen abandoning an object on the ground near the chain-link fence.
Qwen2.5-VL-3B correct
a person enters a restricted area
Qwen2.5-VL-7B probes — asked about Loitering (never happens) → said no · asked about Intrusion (does happen) → said no · time of day: Dawn (gt Dawn) ✓
Intrusion/C00_008_0002
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:17
JoyAI-VL-8B
A person in red clothing is walking along the road in the monitored area.
Qwen2.5-VL-7B correct
A person in a red jacket walks along the sidewalk and then bends down, possibly to pick something up.
Cosmos-Reason1-7B wrong class
A person in a red jacket is seen walking along a fenced pathway, then suddenly stops and bends over, as if dropping something near the ground.
Qwen2.5-VL-3B correct
a person enters a restricted area
Qwen2.5-VL-7B probes — asked about Loitering (never happens) → said no · asked about Intrusion (does happen) → said no · time of day: Day (gt Day) ✓
Loitering/C00_001_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:04
JoyAI-VL-8B
A person in a red top is walking along the sidewalk, which is a normal pedestrian activity.
Qwen2.5-VL-7B correct
A person in a red coat stands on the sidewalk for an extended period.
Cosmos-Reason1-7B wrong class
A person in a red jacket and dark pants is seen walking along a sidewalk, then suddenly stops and drops an object onto the ground.
Qwen2.5-VL-3B wrong class
a person leaves a red coat on the sidewalk
Qwen2.5-VL-7B probes — asked about Intrusion (never happens) → said no · asked about Loitering (does happen) → said no · time of day: Day (gt Day) ✓
Loitering/C00_004_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:50
JoyAI-VL-8B
A person fell down on the sidewalk.
Qwen2.5-VL-7B correct
A person walks slowly along the sidewalk, lingering for an extended period.
Cosmos-Reason1-7B wrong class
A person in a light-colored jacket and dark pants is seen walking along a fenced pathway, gradually moving out of view until they are no longer visible in the frame.
Qwen2.5-VL-3B wrong class
a person leaves a car
Qwen2.5-VL-7B probes — asked about Intrusion (never happens) → said no · asked about Loitering (does happen) → said no · time of day: Day (gt Day) ✓
Violence/C00_015_0003
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:35
JoyAI-VL-8B
A person is seen falling from the roof of a traditional building in the courtyard.
JoyAI-VL-8B
Alert: a person has fallen down in the courtyard.
Qwen2.5-VL-7B correct
A man and a woman are seen fighting in the courtyard.
Cosmos-Reason1-7B correct
Two individuals are engaged in a physical altercation, with one person pushing another, causing them to fall to the ground.
Qwen2.5-VL-3B wrong class
a person enters a restricted area
Qwen2.5-VL-7B probes — asked about Abandonment (never happens) → said no · asked about Violence (does happen) → said no · time of day: Dawn (gt Dawn) ✓
Violence/C00_020_0001
ground truth
JoyAI-VL-8B
Qwen2.5-VL-7B
Cosmos-Reason1-7B
Qwen2.5-VL-3B
00:00:0000:07:11
JoyAI-VL-8B
Abandonment of an object detected: a black object was left on the rooftop and is no longer present.
Qwen2.5-VL-7B correct
Two individuals engage in a physical altercation in the courtyard.
Cosmos-Reason1-7B correct
Two individuals are engaged in a physical altercation, with one person pushing another, causing them to stumble backward.
Qwen2.5-VL-3B wrong class
a person enters a restricted area
Qwen2.5-VL-7B probes — asked about Abandonment (never happens) → said no · asked about Violence (does happen) → said no · time of day: Dawn (gt Dawn) ✓