Real-World Video Understanding David Luan, co-founder
Dextro is a computer vision company whose APIs help companies with lots of video data
to categorize, analyze, and search through it.
From each video, our system extracts:
Brands Objects Scenes
NIKE
CAR
BEACH SCENE
CAR
NIKE
…and turns them into a whole-video summary
SALIENCE GRAPHMusicians
Saxophones
People
as well as a simple timeline.
{
"request_id": "1424722884.021RSBR1LQXE",
"detections":[
{
"id": 2,
"name": "Skyline",
"salience": 0.9,
"thumbnail": "https://api.dextro.co/sample_video_thumbnails/1424722884.021RSBR1LQXE_2844.jpg",
"instance_occurrences": [
[
9.56,
10.52
],
[
11.04,
11.48
],
...
[
75.04,
84.52
]
]
},
],
...
}
DEMO
stream.dextro.co
DISCOVERY CURATION
AUDIENCE
Hand-tuned
"two young girls are playing with lego toy."
ILSVRC 2012
The allure of doing everything:
Medical
Satellite Defect Analysis
Multispectral
Medical
Satellite
Defects
UGCStock Photo
vs
NewsUGC
Entertainment
vs
Stock photos, Google image search
Everything else
ICONIC REAL-WORLD
MOVING BEYOND TAGGING
Useful in stock media context
Sunny, road, trees, grass, green, highway
TAXONOMY
IAB Tier 1 or 2
General UGC taxonomy Custom partner taxonomy
VIDEO-SPECIFIC MODELS
Motion cues are important.