DeepSense or Facial Recognition

Gregory William Keller said:

Has anyone seen these Facial Recognition cameras in real life yet? It is interesting that their FR lineup is 2MP based. The google seems only to know of the press releases. Having a camera with fewer but crisper image frames seems like a big win for FR.

www8.dahuasecurity.com/products/All-Products/Artificial-Intelligence/AI-IPC/Face-Recognition


@EMPIRETECANDY can maybe track a couple down?

If not these, any other recommendations for getting to the ID level of DORI in good light (think tradeshow floor)

Click to expand...

• Low recognition rate: The camera failed to recognize ~30%+ of subjects whose faces were detected, due to multiple issues, with angle of incidence, subjects looking down/away, and low light causing the majority of failures.
• Significant confidence reductions when looking down/away: In our tests, camera confidence decreased by at least ~8-10% when subjects looked down or away from the camera, making it the biggest issue in reliable capture.
• Significant WDR/low light confidence reductions: Additionally, confidence dropped at least 4-5% under more difficult lighting conditions, including simple WDR scenes, 2-5 lux low light, or dark scenes with IR on.
• Reduced or no recognition under 1 lux (without IR): Below 1 lux, recognition became spotty, with the majority of subjects missed. At ~0.5 lux and below, detection did not work at all.
• Multiple instances of mistaken recognition: The camera mistook one person for another 4 times over two weeks of testing in multiple scenarios, using a total database of about 30 people in a small office. In higher volume installations, accuracy is likely to decrease, as additional subjects increase the likelihood that two people will look alike.
• Frequent incorrect demographic information: Demographic information of subjects was frequently wrong, with male/female, glasses/no glasses, beard/no beard, and other factors incorrect much of the time.

john-ipvm said:

We tested the Dahua 8242 face recognition camerashown on that webpage and it had various problems.

From our summary:
The Dahua DH-IPC-HF8242FN-FR had multiple issues making it more difficult to recommend for larger/more widespread use.

• Low recognition rate: The camera failed to recognize ~30%+ of subjects whose faces were detected, due to multiple issues, with angle of incidence, subjects looking down/away, and low light causing the majority of failures.
• Significant confidence reductions when looking down/away: In our tests, camera confidence decreased by at least ~8-10% when subjects looked down or away from the camera, making it the biggest issue in reliable capture.
• Significant WDR/low light confidence reductions: Additionally, confidence dropped at least 4-5% under more difficult lighting conditions, including simple WDR scenes, 2-5 lux low light, or dark scenes with IR on.
• Reduced or no recognition under 1 lux (without IR): Below 1 lux, recognition became spotty, with the majority of subjects missed. At ~0.5 lux and below, detection did not work at all.
• Multiple instances of mistaken recognition: The camera mistook one person for another 4 times over two weeks of testing in multiple scenarios, using a total database of about 30 people in a small office. In higher volume installations, accuracy is likely to decrease, as additional subjects increase the likelihood that two people will look alike.
• Frequent incorrect demographic information: Demographic information of subjects was frequently wrong, with male/female, glasses/no glasses, beard/no beard, and other factors incorrect much of the time.

Click to expand...

john-ipvm said:

Yes, the Chinese government has blocked IPVM. They give no notice nor reason nor appeal. We join the BBC, WSJ, NY Times, etc. as publications blocked inside China.

Click to expand...

Gregory William Keller said:

I'm still curious why they picked 2MP, maybe just because the onboard or NVR processing was limited.

Click to expand...