Please fix this urgently!!!
DeepStack models make a very inaccurate object detection.
A tree leaf is detected as a person 71%
Wow!!!
Object detection - Dog tagged as Person
Re: Object detection - Dog tagged as Person
- Attachments
-
- hoja como persona.jpg (41.17 KiB) Viewed 6349 times
-
- hoja como persona 1.jpg (48.98 KiB) Viewed 6349 times
Re: Object detection - Dog tagged as Person
Hi, I'm afraid we can't as this is a user forum with the odd moderator post. I must admit that I stopped using DS for that reason, and also that it black screened my previous pc. Must be time to try it again
Forum Moderator.
Problem ? Ask and we will try to assist, but please check the Help file.
Problem ? Ask and we will try to assist, but please check the Help file.
Re: Object detection - Dog tagged as Person
- "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
- If you know what your after, you'll recognize it when you see it.
Re: Object detection - Dog tagged as Person
I note that the OP's camera appears to be zoomed in to the grass making that leaf quite large. That could be a problem for DS.
When it happened to me there was a house in the background, and dancing tree shadows on my lawn. DS was very sure that was a person. It really is time for me to try DS again. That was the very first version that integrated fully with BI5, and it kept black screening my server with a crash an hour later with no logs at all. Sam Varghese tried to assist, but I eventually deleted DS as it was my main BI server
When it happened to me there was a house in the background, and dancing tree shadows on my lawn. DS was very sure that was a person. It really is time for me to try DS again. That was the very first version that integrated fully with BI5, and it kept black screening my server with a crash an hour later with no logs at all. Sam Varghese tried to assist, but I eventually deleted DS as it was my main BI server
Forum Moderator.
Problem ? Ask and we will try to assist, but please check the Help file.
Problem ? Ask and we will try to assist, but please check the Help file.
Re: Object detection - Dog tagged as Person
When it happened to me there was a house in the background, and dancing tree shadows on my lawn. DS was very sure that was a person.
Re: Object detection - Dog tagged as Person
.
.
I'm not working with zoom. !!!
Take a look to real image capture from my camera in the link below: https://imgur.com/a/ZI7bd8p
.
.
.
- Attachments
-
- hojas como persona & persona.jpg (183.64 KiB) Viewed 6274 times
Re: Object detection - Dog tagged as Person
Evaluate results setting AI confirmation percent to 60. Is that last image a JPEG? An alert image? Or is that a capture using Test & Tune during playback?
- "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
- If you know what your after, you'll recognize it when you see it.
Re: Object detection - Dog tagged as Person
Thought so. Wrong tool for the evaluation. Not your fault, it's not well documented. I recommend reading Canceled Alert with 91% deepstack in order to get your arms around DeepStack analysis.
There are two ways to view a DeepStack analysis: Review the DAT or Analyze the BVR. Each deliver different information. Use testing and tuning analyze with DS to review recordings and see what DS identifies - especially useful for face recognition or custom models. The DAT is a log of sorts so that one can review when DS began and what it found, so you can fine tune timing of camera's or streams or buffers and the like. Running T&T DS analysis on a recording doesn't have network/bandwith/stream issues - it's fast delivery to DS. So it doesn't really represent the same conditions.
There are two ways to view a DeepStack analysis: Review the DAT or Analyze the BVR. Each deliver different information. Use testing and tuning analyze with DS to review recordings and see what DS identifies - especially useful for face recognition or custom models. The DAT is a log of sorts so that one can review when DS began and what it found, so you can fine tune timing of camera's or streams or buffers and the like. Running T&T DS analysis on a recording doesn't have network/bandwith/stream issues - it's fast delivery to DS. So it doesn't really represent the same conditions.
- "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
- If you know what your after, you'll recognize it when you see it.
Re: Object detection - Dog tagged as Person
Let me make a case to support my point with the illustration below (click to enlarge).
Percent Confidence Comparison - Analyze with Deepstack vs Deepstack tab (BVR vs Real Time) Streaming quality, dropped frames, key frame rate and interval, network bandwidth, CPU, AI settings and quality... all that impacts the success of real time object recognition. A ton of stuff.
About 6 months ago I was screaming bloody murder trying to understand Deepstack. The pivotal event that changed everything for me, was getting the cameras to be Plain Jane settings per step 2 of the Camera Connector - IP Config Dialog in the self help. Once I did that, it was easy to refine and tune Deepstack because the stream was pretty generic from a data perspective.
In my view, Blue Iris is as complicated as Photoshop - if not more so. There are easy bits and there are "WTF" bits. It takes time and most of all managed expectations. You don't need perfect, you need it to do the job. That's how I've come to see things.
That said, if you want help or clarification, it's here... if you want.
Pro tip: Status window -> Deepstack tab -> Ctrl-Dbl Click on clip. The clip will open and pause at the triggering event, and the Deepstack tab will populate. Enable show motion rectangles in the DS tab.
Percent Confidence Comparison - Analyze with Deepstack vs Deepstack tab (BVR vs Real Time) Streaming quality, dropped frames, key frame rate and interval, network bandwidth, CPU, AI settings and quality... all that impacts the success of real time object recognition. A ton of stuff.
About 6 months ago I was screaming bloody murder trying to understand Deepstack. The pivotal event that changed everything for me, was getting the cameras to be Plain Jane settings per step 2 of the Camera Connector - IP Config Dialog in the self help. Once I did that, it was easy to refine and tune Deepstack because the stream was pretty generic from a data perspective.
In my view, Blue Iris is as complicated as Photoshop - if not more so. There are easy bits and there are "WTF" bits. It takes time and most of all managed expectations. You don't need perfect, you need it to do the job. That's how I've come to see things.
That said, if you want help or clarification, it's here... if you want.