5 ms·

I'll be a believer when it can handle a Thomas Crown scenario - multiple people in the same outfit crisscrossing around each other. While a bowler hat flashmob

I'll be a believer when it can handle a Thomas Crown scenario - multiple people in the same outfit crisscrossing around each other. While a bowler hat flashmob is unlikely, a group of similar dark jackets etc isn't.

jerf 10y ago
Actually, you can't forget the possibility that someone will attack it, for profit or even just the lulz. If you have the idea I can pretty much promise someone's going to try it long before there's a 1000 of these stores.
- wongarsu 10y ago
  People are attacking regular stores' checkout systems already. We call those people shoplifters. And our normal method of dealing with them (have a human that catches a few of them to deter the more law-abiding citizens) will work equally well in a checkout-less store. The big new challenge is telling the difference between a shoplifter and a normal software error. But that is only relevent for the most sophisticated shoplifting ideas.
  - illumin8 10y ago
    Exactly. And, since you have to scan your smartphone before entering, they already have your name, address, credit card details, etc., so shoplifting would be extremely risky. Most teenage shoplifters might be able to outrun a security guard that saw them leaving with unpaid items, but what good does it do you when the store already has all your personal details and can simply send the police to pick you up at home?
    - chronicx0 10y ago
      What if 4 shoppers are wearing the same black Google hoodie, noogler hat, and walk closely as a group? Good luck on phone or visual localization and/or tracking.
  - walrus01 10y ago
    steak = 4011
dangero 10y ago
This scenario is the same as any other machine learning environment: The output is prediction with a % confidence score. This means the threshold can be set such that worst case certain people need to be checked out by hand. Fallback to standard checkout means they still cut down quite a bit on the needed human interaction and they can set KPIs around improving those stats.
- antognini 10y ago
  While it's sometimes okay to interpret the output of a model as a confidence, you do so at your own peril because that's not really what the model is optimizing. There are often cases where the model outputs a very high "confidence", but the model uncertainty is (or should be) very high. This is generally the case anytime you give the model data that is very far (in some sense) from the data that it's been trained on.
Splines 10y ago
Protip: Avoid going to the closest Amazon Go store after a Seahawks game. Unless you're wearing Broncos gear.

jerf 10y ago

Actually, you can't forget the possibility that someone will attack it, for profit or even just the lulz. If you have the idea I can pretty much promise someone's going to try it long before there's a 1000 of these stores.

wongarsu 10y ago
People are attacking regular stores' checkout systems already. We call those people shoplifters. And our normal method of dealing with them (have a human that catches a few of them to deter the more law-abiding citizens) will work equally well in a checkout-less store. The big new challenge is telling the difference between a shoplifter and a normal software error. But that is only relevent for the most sophisticated shoplifting ideas.
- illumin8 10y ago
  Exactly. And, since you have to scan your smartphone before entering, they already have your name, address, credit card details, etc., so shoplifting would be extremely risky. Most teenage shoplifters might be able to outrun a security guard that saw them leaving with unpaid items, but what good does it do you when the store already has all your personal details and can simply send the police to pick you up at home?
  - chronicx0 10y ago
    What if 4 shoppers are wearing the same black Google hoodie, noogler hat, and walk closely as a group? Good luck on phone or visual localization and/or tracking.
- walrus01 10y ago
  steak = 4011

dangero 10y ago

This scenario is the same as any other machine learning environment: The output is prediction with a % confidence score. This means the threshold can be set such that worst case certain people need to be checked out by hand. Fallback to standard checkout means they still cut down quite a bit on the needed human interaction and they can set KPIs around improving those stats.

antognini 10y ago
While it's sometimes okay to interpret the output of a model as a confidence, you do so at your own peril because that's not really what the model is optimizing. There are often cases where the model outputs a very high "confidence", but the model uncertainty is (or should be) very high. This is generally the case anytime you give the model data that is very far (in some sense) from the data that it's been trained on.

Splines 10y ago

Protip: Avoid going to the closest Amazon Go store after a Seahawks game. Unless you're wearing Broncos gear.