So if you send main stream images to DS it will have to work harder to downsize the image before it tries to detect the object. The only benefit is the image that BI save for the alert will be at the higher resolution.
IP Cam Talk Custom Community DeepStack Model
- Thread starter MikeLud1
- Start date
sebastiantombs
Known around here
I've just found, empirically with experience over time, that using the main stream results in lower detection rates and timeouts at times. Using the sub stream works well for me on all the cameras I have DS running on. The jury is still out a little on that, though. I just added a VTO-220 on a test burn in. The sub stream resolution is so low that it's basically not useable for me so I've only enabled main stream. The main stream is relatively low resolution as well, but so far it's work well. Once it's installed in its' "real" location, rather than sitting on a shelf in my office, I'll find out more in that regard.
I just did a test setting my sub stream to VGA 640 x 480 and the accuracy is the same as before sending 1080P images the speed dropped from ~75ms to ~45ms.
sebastiantombs
Known around here
With the custom models I made people with CPU version of DeepStack are almost see GPU speeds.Thanks for doing this @MikeLud1
I'm not using a custom file as yet, but this may be the push that I need.
I'm using a I7-7700 and won't be adding a GPU, so this may work!
sebastiantombs
Known around here
@sebastiantombs Did some testing this afternoon. Turning off "Use main stream" was the fix for my issues with running the Combined and Dark models at the same time. Thanks!
For people that are using the CPU version of DeepStack please post you detections time before the new custom models and after. I am curious to see how much the new models are helping with the detection times. From some posts the speed increase looks to be about 7 to 10 times faster.
My wife just arrived. It's dark outside and none of the 5 cameras along the driveway identified her car. I'm running Combined and Dark with 70% confidence. I ran several of the clips using "Analyze with Deepstack" under "Testing & Tuning" and every clip showed multiple confirmations of a car in the 90% range. Any idea why it didn't confirm and trigger?
sebastiantombs
Known around here
I'd drop confidence to 40% on those cameras, especially for night detection even with dark.pt in the mix. If you look at the screen shots I posted, the second screen shot is night parameters that I use, and work in pretty difficult circumstances, work well here.
Perhaps but the testing showed multiple hits over 80%. Just trying to understand what happened.
sebastiantombs
Known around here
You need to understand that "testing" with DS is a far cry from what happens during normal detection. If DS worked as well, real time, as it does when "testing and tuning" we wouldn't have anything to talk about 
The "Analyze with Deepstack" under "Testing & Tuning" will ALWAYS perform better than live as it is after the fact and should not be used as an analysis tool to try to figure out why it didn't see a car or person. It should only be used to see what DeepStack can find in that clip, like "hmm I wonder if DeepStack can find a toothbrush" and then walk around with a toothbrush and have it identify it. I can run this on a camera not using Deepstack and it will show EVERYTHING that Deepstack has in its objects to find that it sees in the clip.
You need to use the .DAT file analysis that shows how DeepStack responded and behaved live. You have to tell it under AI for the camera to check the Deepstack analysis option. Only then can you start to figure out why it missed these events.
I agree, drop the % confidence - do you want it to find it or not? Having it at 70% and missing doesn't do you any good. Drop it down until you start getting false. I have all of mine low and never get a false.
You need to use the .DAT file analysis that shows how DeepStack responded and behaved live. You have to tell it under AI for the camera to check the Deepstack analysis option. Only then can you start to figure out why it missed these events.
I agree, drop the % confidence - do you want it to find it or not? Having it at 70% and missing doesn't do you any good. Drop it down until you start getting false. I have all of mine low and never get a false.
sebastiantombs
Known around here
Just check the box labeled "save deep stack analysis details" in the AI config of the camera(s). Use file explorer to find the file in the directory you're saving alerts in. Drag and drop into the deepstack tab of the BI status console. Be aware that it can add up to a lot of files very quickly.
Open up the BI status page and then there is a DeepStack tab. Open that tab and then minimize that screen.
Then open up the cancelled alert. Once it starts playing, pause it and then open up the minimized screen and it will show all the details on when and what it looked for.
Based on looking at the .DAT file, it looks like my + Real-time images setting was too low. I had it set to 10 and every 500ms. I think it finished checking before the car was in the frame.
That will do it. One way to force it to run thru all the images is to put an object in the to cancel box - like a banana - it will then run thru each image. But depending on your computer, use this sparingly as it will start to drive up CPU % and make times. But if you have one problematic field of view, this works very well!
I use it on my cams at night to ID cars that were getting missed due to headlight shine on the pavement triggering it early.
sebastiantombs
Known around here
Multiply the milliseconds by how many images are analyzed, then look at the pre-trigger time. You need to give DS enough images with the target in frame.