Again, that is why I said "No disrespect". I get it, I actually LIKE being a guinea pig (to a certain extent). It is just with .9.3 .9.4 is seemed like most people had DeepStack 100 within seconds on the upgrade. My emphisis should have been on the "Are they doing something different than us?". Up until .3 and .4 I have been thrilled beyond belief with the progress. Guess I got spoiled.
Again, that is why I said "No disrespect". I get it, I actually LIKE being a guinea pig (to a certain extent). It is just with .9.3 .9.4 is seemed like most people had DeepStack 100 within seconds on the upgrade. My emphisis should have been on the "Are they doing something different than us?". Up until .3 and .4 I have been thrilled beyond belief with the progress. Guess I got spoiled.
I'd like to know if the errors are constrained to the custom portion of the analysis. A couple of screenshots would be helpful too
Were you able to download 5492?
Yes, got it thanks!
Here is the analysis screen when analyzing a clip with DeepStack Server Error 100. Not sure if that is what you want of if it even helps. But want to be part of the solution!
That's exactly what I was looking for. Notice how all your error msgs are in the custom portion of each frame's analysis? If this is the situation for all of your cameras, then if you stop using your custom model(s), you shouldn't get any more error 100s. Also notice how the asterisk appears in your second frame at T+1001--that tells me the alert was confirmed at that point, but frame analysis continued (maybe you specified extra frames for analysis at, it looks like 1 second intervals), or you have something in the To Cancel box of your trigger requirements in the AI tab. This alert was probably saved because it was confirmed early, but too often, an early error 100 will prevent the alert from proceeding to confirmation and then you'll have missed capturing an event. Thanks for uploading; that's very helpful. I stopped using custom profiles and have had no error 100s as of yet. I think I'll wait a while before reintroducing dark.pt.
Also notice you have Car,car,Truck,truck in your To Confirm box....BI/Deepstack will do it's thing with car,truck,person,people, etc without having to provide a second capitalized word. Give it a shot and you'll still get the appropriate custom reports.
^^^ So what you're saying is that custom models feature doesn't actually work, or work properly. It doesn't explain the "can't find the file" problems either. In any case, it just isn't working as advertised. Given that 5.4.8.2 is working with custom models and always finds the file I think I'll stick with that unless and until another reliable version comes along.
Almost every instance of errors when people post logs (and looking at my own), is it seems to correlate with time exceeding 5,000ms
And yeah I get how some can be upset about us being beta testers, but there are literally so many different configurations - I bet not one of us here is running an identical setup of hardware, software, Windows version, etc.
I think the gains we get out of the program far exceed the setbacks. I was tickled pink with my BI setup about a year ago and left it as as and didn't figure there was anything he could do to improve it and then he added the substreams game changer, ability to record substream until mainstream and DeepStack - and those are just the ones that I found useful, but dozens of other bonuses have been added that others have found useful and updated with.
^^^ So what you're saying is that custom models feature doesn't actually work, or work properly. It doesn't explain the "can't find the file" problems either. In any case, it just isn't working as advertised. Given that 5.4.8.2 is working with custom models and always finds the file I think I'll stick with that unless and until another reliable version comes along.
I'm speculating, but when a trigger occurs, a thumbnail is created while BI/DS sort out whether or not to confirm. In some cases the thumbnail remains even though there's no saved file associated with it--I think you've experienced this--the thumbnail is marked with error 100--I've not experienced this. In other cases which I've experienced, the thumbnail goes away when the alert is cancelled under normal operation or when the alert is incorrectly cancelled because of an error 100. I've monitored this and can access the .dat file to confirm my suspicion...it looks much like the graphic kc8tmv posted above. I can duplicate this by manually (externally) triggering a camera with no motion present. The thumbnail is created as a snapshot but disappears when there's no actual object in motion.
As of v5.4.9.3 installed 1-July-2021 I am now getting the error 100 a lot in the log... Upgraded to 5.4.9.4 today and still getting the error 100. I noticed it a few versions back but it had stopped.
I think the Error 100 is a timeout issue with DS.
Almost every instance of errors when people post logs (and looking at my own), is it seems to correlate with time exceeding 5,000ms
I don't know of any. But you can make DS timeout quickly by setting 0 additional real time images and leaving the To Cancel box empty. To give DS the longest opportunity to analyze, increase real time images very high...of course this would be at the cost of a quick alert confirmation.
As of right now, we do not have that ability with the integrated DS/BI.
This is one of the many granular abilities that folks have if they are using a 3rd party add-on like AI tools or one of the many other programs using DS as the AI engine. I believe AI tools allows one to set up a buffer queue so that if DS gets slammed with requests, it can work off that queue, and I think at the moment it looks like BI simply times out.
It doesn't mean it won't come to BI, but we can see with all the issues folks have, that BI needs to let stuff out in pieces to see how the systems work.
The next "big" update will allow us to specify a time shorter than one second to send photos to DS - while that can help those tricky field of views where an object is not in the field of view very long, on the flip side it will then mean that one could send double or triple the number of photos in the same time frame, so it could create an even bigger bottleneck on less capable machines or having too many images thrown at it at once.
As of v5.4.9.3 installed 1-July-2021 I am now getting the error 100 a lot in the log... Upgraded to 5.4.9.4 today and still getting the error 100. I noticed it a few versions back but it had stopped.
So, you're again saying custom models don't work in 5.4.9.x which is a little disappointing. I think I'll try reducing the number of images down to three to see if that helps a bit and keeps things from timing out when I try the next 5.4.9.x iteration. Maybe I won't have to revert back to 5.4.8.2 again if that pans out although without using dark.pt it's pretty useless at night for my purposes.
I just found out I need to make an addition to my detected objects lists. A few neighbors have horses, we're in a pretty rural area. A couple just went running back and forth. DS did detect the "retreating view" as a person, but I'd rather have it identified as a horse
Another note. I got a response to my original email to BI support about 5.4.9.1 concerning the errors and file not found problems. The response said it needs a key frame ratio of .5. All my cameras are at a ratio of 1, 15FPS/15 iframe. They also want to know if it happens in 5.4.9.4, which it does. They additionally asked for record tab settings to try to duplicate the problem. I'll grab some screen caps later today and email them back.
There appears to be an issue with custom models to the point where it makes DS useless. However, I removed the dark.pt I was using yesterday and BI triggered an alert as a person was walking past one of my cameras last night so the built-in model seems to be better.
As I've said, models work fine in 5.4.8.2 for me. I tried shutting them off in 5.4.9.x versions and still had problems and the standard model is pretty much useless for me at night. I know that's a function of target size, contrast and lighting, but I can't change those parameters without a tremendous amount of work so the .9.x series is off the table for me until models work in it. I'll also say, yet again, that if it worked in real time like it works in tuning it would be the bees knees.
WOW - up to this point the BI recommendation has been a KEY of 1.00 - lower KEY rates below 1.00 has been why Reolinks have trouble with motion detection in BI.
I have found the dark model to work better than the built -in. Even in daytime, my dark % is always higher than the built-in.
My issue has been when running an update that allows custom models, the rest of the system gets unstable.
But I may be on to something that I am testing and will report back later if it pans out.
A KEY of 0.5 means it sends a frame every 2 seconds instead of 1 second, so FPS/iframe of 20/40 or 10/20 is both a KEY of 0.5, just as FPS of 30 and iframe of 60 is.
