Lighthouse Deep Learning Camera Tested

Avatar
Rob Kilpatrick
Published Dec 07, 2017 15:41 PM

A Silicon Valley startup, Lighthouse [link no longer available], with a Stanford PhD CTO, has released a deep learning AI camera with 3D sensors for just $300.

The company claims "False alerts, gone forever. Really" and more.

We bought one and tested their claims. Inside, we share our findings on:

  • Adult / Child / Animal Detection
  • Facial Recognition
  • Voice Search / Voice AlertSsetup
  • Home / Away Detection
  • Mobile App Usability 
  • And More

Summary

Overall, Lighthouse performed well in our tests, with minimal false alerts plus object classification and facial recognition above and beyond what other consumer/DIY cloud offerings have claimed and failed to deliver, such as Nest IQ or Netatmo Welcome

However, users should beware of gesture detection claims which simply do not work (running, jumping) as well as potential failed recognitions which may create nuisance or missed alerts. 

Finally, the Lighthouse app is fairly limited compared to many cloud cameras, with limited options for search, playback, and export.

Market Impact

The biggest impact of Lighthouse's performance is to demonstrate that a team can deliver an effective / advanced deep learning camera. Because it is a startup aimed at the consumer space, it will not have any immediate impact on the broader / larger commercial market where we believe their features would be more highly valued.

In the consumer market, Lighthouse does have clear technical differentiation, but at a price point of $300 + $100 per year required subscription, it will limit itself to higher disposable income users. On the other hand, there are certainly enough well to do individuals who can afford this that the company may grow quickly tapping into that niche.

Key Findings

In our testing there were several positive results from Lighthouse, all of which are rare among consumer cameras and even many commercial offerings:

  • Object detection reliable: Lighthouse was consistently able to differentiate between adults, children, and pets.
  • Facial recognition reliable (with caveats): Known people were generally reliably recognized, but some issues with misidentification or attempting to recognize subjects not facing the camera were present.
  • Natural language processing / voice search: Search and alert setup using natural language, including speech to text, worked well, even for complex multi-criteria searches.
  • Reliable geofencing: Lighthouse automatically enabled/disabled alerts based on users' locations, with no manual configuration or status changes required.

There were several shortcomings, as well:

  • "Gesture" detection works poorly: While Lighthouse claims users may search for/be alerted upon specific actions like running, jumping, etc., in our tests these gestures were not recognized, with only "waving" working properly.
  • Recognition attempted on turned/obscured subjects: Lighthouse frequently attempted to recognize users which were turned away from the camera or otherwise obscured, resulting in "unknown person" notifications. This causes problems for users who want alerts on unknown persons since many will be nuisance notifications.
  • Occasional misclassifications: Inanimate objects were very rarely classified as people or pets, such as a sliding door being classified a person or a chair classified as a dog. Only a few of these issues were seen in our tests (<5). 
  • Event list limited: Users cannot share/export video directly from clips on the app's event list, meaning they must navigate to the specific event they wish to share by searching for it, which may be frustrating or difficult for events older than the recent past.
  • No playback controls: There is no way to pause, rewind, or fast forward video clips. Instead, clips simply play start to finish and loop, similar to an animated .gif.

Pricing 

Lighthouse can be purchased directly from the manufacturer for ~$300 USD, though preorders are currently sold out. Lighthouse also requires a $100/year subscription plan, not optional.

This pricing is significantly higher than most DIY/consumer models, generally <$200, though similar to the Nest Cam IQ, $300 online.

Physical Overview

Lighthouse is significantly larger than most consumer/DIY cameras, using a cylinder/tube form factor about 8" tall, shown below compared to the Nest Cam IQ:

We review the physical features of the camera in this video:

Lighthouse Marketing vs How It Works

Lighthouse marketing somestimes does not accurately represent the system's actual use. While it works, it is not as a clearly or directly described in the app.

For example, they market that they notify when it 'looks like someone new is here' but what this is, in reality, is just an 'unknown' person who in practice is often a known person just not matched, as shown below:

Also for the marketing claim of 'dog walker has arrived', Lighthouse will send an arrival notification for users whose phones are registered in the system but that notification does not include a picture. However, if the dog walker's face is registered, whenever the dog walker is recognized while notifications are enabled, a notification will be sent with their picture.

Object Recognition: Adult/Child/Pet Works Reliably

Over the course of a week of testing, in an open office as well as a typical home setting, Lighthouse's object recognition worked well, classifying and highlighting known/unknown adults, along with children, and pets with few incorrect classifications.

Lighthouse highlights the detected object in a different color depending on what it is, shown below:

Object Classification Errors

Throughout testing, we saw few errors in recognizing objects, 10-20 incorrect classifications out of 500+ events.

Most commonly, the camera merged other objects with moving people, such as the chair shown next to the subject below. Pets were occasionally merged with people, as well.

Additionally, sometimes moving objects were recognized as people, such as the sliding door shown here:

Facial Recognition

Lighthouse recognizes known subjects using facial recognition, which worked fairly well in our tests, with few missed or false identifications (an unknown person recognized as known or vice versa) or mis-identifications (one known person recognized as another).

To improve recognition, users may tag images which Lighthouse suggests may be known users as correct or incorrect, shown below. Note that these images do not trigger alerts for the suggested person.

Occasional Missed Identification

The biggest issue in missed identification was Lighthouse attempting to recognize subjects which were turned away from the camera or otherwise obscured. For example, in the clip below the subject is shown as unknown because the camera cannot properly see his face. Issues such as these may contribute to users losing trust in the system's alert accuracy and potentially ignoring unknown person alerts.

Voice Search

Lighthouse includes natural language processing, which allows users to search for events or set up alerts using typical phrases instead of having to select multiple criteria for search through other controls.

For example in the video below, we use voice search to search for clips of a specific person (based on facial recognition) in a specific timeframe (yesterday, 9AM-12PM).

And in this example, we tell Lighthouse to alert the user (which they call "pings") when it sees a specific person.

Natural Language Limitations

Natural language is limited in what it may show, only working with specific objects (people, children, kids, pets, dogs, cats) and time ranges. Searching/alerting on actions other than running, jumping, or waving do not work, nor do colors.

Gesture/Action Recognition Limitation

In our tests, the only gesture recognition which worked properly was waving, shown below:

Lighthouse's use case for waving is for a child to trigger a notification to their parents who can they respond / see and talk tot he child.

Searching for running simply showed all events due to Lighthouse perceiving all movement as running:

Lighthouse says they plan to improve this feature over time and add more gesture types, but did not give a timeline.

Reliable Automatic Arm/Disarm Geofencing

By default, Lighthouse automatically armed and disarmed alerts based on users' mobile phone location, with alerts sent for all detections when all users are away. The only setup required for this feature is setting the camera's location, but no other configuration or manual status changes are required or even available. This geofencing may also be used in searches/alerts, so users may search "Alert me on unknown people when John is home but I am not" as one example.

 

Test Parameters

The following firmware version was used in this test:

  • Version 1.5.387
Comments are shown for subscribers only. Login or Join