Guys- Great job on the revised recognition engine and the processor use it saves...but I have noticed that it's not picking up events as well as it used to.
Link to an example: https://drive.google.com/open?id=0B4PzC ... mRuNjR1WUE
Not only did Sighthound not flag this as "people", it didn't flag it all, even as "unknown". This camera and 2 others missed this, but Sighthound did correctly flag this person and doggahs as "people" on the 4th camera as she walked.
Granted, this was right after the "flip", going from night mode to day mode, and the light levels are relatively low, so the image is grainy, which I'm sure gives the recognition engine trouble.
Any input is appreciated.
Failure to recognize events
Moderator: Staff
4 posts
• Page 1 of 1
-
Kapitoshka - Posts: 3
- Joined: Wed Mar 22, 2017 10:26 pm
Re: Failure to recognize events
I have 2 different IP cams (3 altogether) and one USB HD cam. I noticed that version 4 misses a lot of people recognition. I'm not talking about the tricky ones. Just like a previous poster. People in the clear sight are not recognised or/and reported. I repeatedly send proof to support, but response is there is nothing they can do at the moment. If I'm not happy with version 4 I'm more than welcome to downgrade to v3. Well, not good enough for something that cost that much money! It is becoming kind of half useless to us as v3 CPU optimisation so poor that NUC i5 run at 100% CPU usage just on 4 cameras and becoming useless. V 4 is much better, it uses 50% - 60% CPU, but misses events. What do you do with product that either not fit for purpose or not performing as advertised, yet support sees the problem and not even hints on ETA for fix?
To assist support further I confirmed that it happens on different camera types, in different environments (e.g. indoors and outdoors), it does not discriminate day/transition/night etc. Well happens a lot anyways, here is one example from today: https://drive.google.com/open?id=0B0SiXvM4aZFvazlfWE5SY1ZSTDA
To assist support further I confirmed that it happens on different camera types, in different environments (e.g. indoors and outdoors), it does not discriminate day/transition/night etc. Well happens a lot anyways, here is one example from today: https://drive.google.com/open?id=0B0SiXvM4aZFvazlfWE5SY1ZSTDA
-
Devin - Posts: 719
- Joined: Wed Mar 12, 2014 5:48 pm
Re: Failure to recognize events
Hello All,
Thank you for your feedback on the latest update. We are constantly working to improve our algorithms. Tests on our database of videos and most feedback from users show that 4.X is at worst similar to and often better than 3.X. Work has also not stopped on the algorithms in Sighthound Video, but algorithm development of this sort is a very different task than something like addressing a crashing bug. There isn't a single fix to find and every adjustment has its tradeoffs. That said, there are ways to improve the recognition accuracy in your scenes. Sighthound needs to make a classification decision in a short amount of time so that responses for your rules (alerts, exports, etc…) can be sent. These can also not be taken back so the label cannot be retroactively altered. Accuracy is best if the full object is clearly visible early in its existence.
@jbrannen - You are correct in that the "flip" from night to day mode, as well as grain, can cause problems with Sighthound. This is because Sighthound 'learns' the background of your scene and makes decisions about what a moving object is using this knowledge. If the flip happened shortly before this event it's possible that it was still learning the new background and was not as accurate during this period. Another potential problem in this video is that as the object enters the tree is obscuring it and the mailbox blends quite a bit. I suspect if this person in your video entered the scene from the left the person would have been correctly detected. You may see improved accuracy by shifting the angle of the camera slightly so that the mailbox and tree aren't at the edge of the scene where people enter.
@Kapitoshika - This video and screenshots you emailed in look as though there are objects obstructing the person as they enter, such as a car, gate, stroller or the person is standing with only part of their body in the scene. In the livingroom screenshots you provided it appears many objects are only partially visible at the edge of the scene. There was also a screenshot of a person sitting on a couch with the arm tracked. If an item is motionless in the scene for a period of time it will be learned into the scene. If only part of it moves, such as an arm of an otherwise motionless person, this arm similarly will likely not be identified as a person on its own. If you are completely dissatisfied with the 4.X and paid for a subscription renewal we can certainly issue a refund.
More general tips for getting best recognition can be found in this support article: https://www.sighthound.com/support/tips ... ecognition
Thank you for your feedback on the latest update. We are constantly working to improve our algorithms. Tests on our database of videos and most feedback from users show that 4.X is at worst similar to and often better than 3.X. Work has also not stopped on the algorithms in Sighthound Video, but algorithm development of this sort is a very different task than something like addressing a crashing bug. There isn't a single fix to find and every adjustment has its tradeoffs. That said, there are ways to improve the recognition accuracy in your scenes. Sighthound needs to make a classification decision in a short amount of time so that responses for your rules (alerts, exports, etc…) can be sent. These can also not be taken back so the label cannot be retroactively altered. Accuracy is best if the full object is clearly visible early in its existence.
@jbrannen - You are correct in that the "flip" from night to day mode, as well as grain, can cause problems with Sighthound. This is because Sighthound 'learns' the background of your scene and makes decisions about what a moving object is using this knowledge. If the flip happened shortly before this event it's possible that it was still learning the new background and was not as accurate during this period. Another potential problem in this video is that as the object enters the tree is obscuring it and the mailbox blends quite a bit. I suspect if this person in your video entered the scene from the left the person would have been correctly detected. You may see improved accuracy by shifting the angle of the camera slightly so that the mailbox and tree aren't at the edge of the scene where people enter.
@Kapitoshika - This video and screenshots you emailed in look as though there are objects obstructing the person as they enter, such as a car, gate, stroller or the person is standing with only part of their body in the scene. In the livingroom screenshots you provided it appears many objects are only partially visible at the edge of the scene. There was also a screenshot of a person sitting on a couch with the arm tracked. If an item is motionless in the scene for a period of time it will be learned into the scene. If only part of it moves, such as an arm of an otherwise motionless person, this arm similarly will likely not be identified as a person on its own. If you are completely dissatisfied with the 4.X and paid for a subscription renewal we can certainly issue a refund.
More general tips for getting best recognition can be found in this support article: https://www.sighthound.com/support/tips ... ecognition
-
Kapitoshka - Posts: 3
- Joined: Wed Mar 22, 2017 10:26 pm
Re: Failure to recognize events
Hi Devin,
I appreciate your concise and honest report. As you may be aware that people buy security software to capture “unexpected”. The whole expectation of clear person image is tiny bit unreasonable. Many of your customers buy your product to monitor deliveries, people with foreign objects entering the environment, people entering and exiting cars etc. As I value your hard work, consistent support and great communication I have no intention to ask you for refund. I guess my point is that, if you can twick your product to be more accustomed/agile to people with/behind/aside objects that would be highly beneficial. Whilst I’m happy to assist you with perfecting your product I don’t particularly want to be in situation where in a year time, when I just run out of support not be able to upgrade to the version that actually works for me. Hope you understand our point of view as much as we understand yours. I’m more than happy to take our discussion offline if this is not appropriate channel for such debate.
Kind regards,
Greg
I appreciate your concise and honest report. As you may be aware that people buy security software to capture “unexpected”. The whole expectation of clear person image is tiny bit unreasonable. Many of your customers buy your product to monitor deliveries, people with foreign objects entering the environment, people entering and exiting cars etc. As I value your hard work, consistent support and great communication I have no intention to ask you for refund. I guess my point is that, if you can twick your product to be more accustomed/agile to people with/behind/aside objects that would be highly beneficial. Whilst I’m happy to assist you with perfecting your product I don’t particularly want to be in situation where in a year time, when I just run out of support not be able to upgrade to the version that actually works for me. Hope you understand our point of view as much as we understand yours. I’m more than happy to take our discussion offline if this is not appropriate channel for such debate.
Kind regards,
Greg