NICE Suspect Search ExaminedBy John Honovich, Published Sep 25, 2014, 12:00am EDT
Did you lose your child here?
Now, many might say you were a terrible parent but not NICE Systems. NICE's new Suspect Search [link no longer available] aims to solve this.
In this note, we break down how it works and what the problems with it are.
First, watch their one minute marketing video for the pitch:
Their demo video [link no longer available] shows searching 14 hours in 3 minutes, sort of.
Step #1 is to create an avatar, picking from skirts, skinny jeans, cargo shorts, hair style, body types, coats, etc.
Then NICE returns matches. Scarily, even their marketing video matches are weak:
If you find any correct matches, then you can do a secondary search with them:
However, the results are still poor, as shown by their own marketing demo:
Finally, for the matches you do get, NICE will generate an overlay on a map showing when and where the subject was over time:
[UPDATED] Technical Background
NICE reported the following tech details:
- When analyzing video, NICE pulls and detects people. Those people are then analyzed based on color, texture and other features, enabling people to be compared and contrasted.
- The maximum FoV width to be analyzed is ~10 meters based on their current analysis using CIF resolution. As such, the cameras can have moderately wide FoVs at best.
- For ~100 cameras, this would require an additional server with 12 cores (2 CPUs, 6 cores each) and 32GB RAM.
- This requires using NICE's video management software / system and cannot be 'added on' to 3rd party VMSes.
Doing this accurately is really hard because of:
- Lighting variances - If it is dark or bright or there are shadows, matching accuracy decreases significantly. Of course, these conditions are common in video surveillance.
- Camera positioning variances - Cameras differ in how high they are mounted, how far they are down tilted, side tilted, etc. This creates varying views of people (from behind, from the front, from the side, from above, etc.). Matching accuracy across different views is challenging / poor.
- Crowds - The more people and the more barriers (poles, walls, etc.), the more 'suspects' are obscured / blocked, making it harder for the system to match / recognize them.
Even if the analytics are 'great', these are fundamental problems that cannot be overcome.
There's not many people offering it, certainly not actively today in 2014 (e.g., VideoIQ promoted this back in 2008 but had long stopped claiming this).
There are clearly many large end users who would love to be able to track down a child, a terrorist, a bank robber, Jack Bauer, etc.
As such, we think this is something that will attract interest to NICE, but something that buyers should very carefully test at scale to determine if it can work well enough to be useful in their environments.
New video added:
4 reports cite this report:
Back to Top