I wish to convey my admiration for the state of your reliability monitor; it truly stands out for its impressive performance.
My setup, however, has encountered challenges, particularly following a Blue Iris crash yesterday, which I believe was triggered by my balcony camera freezing, necessitating a restart of the application.
Upon reviewing the reliability monitor, I've noted several Blue Iris crashes, which underline persistent instability issues tied to CPAI and CUDA errors, as illustrated below.
- Real-Mon.jpg (65.21 KiB) Viewed 805 times
Adopting Yolo.NET (DirectML) has, fortunately, led to a crash-free experience in relation to the GPU, puzzling me about the underlying causes of CUDA's instability. DirectML has not crashed since:
- uptime.jpg (12.5 KiB) Viewed 805 times
This experience has solidified my decision to continue using Yolo.NET moving forward.
The switch to .NET has not only stabilised the system but also seems to have slightly improved inference times. When comparing performance, the medium model under .NET appears to either match or slightly outperform CUDA's speed with a small model, which is quite fascinating considering CUDA's hardware specificity. However, faced with CUDA's consistent failure after 11 hours of operation, I find myself with no choice but to adopt DirectML.
An intriguing point I've observed, albeit not definitively linked to .NET, is a detectable inconsistency in CPAI's object detection capabilities, as highlighted below:
- entry-nothing-found.jpg (100.91 KiB) Viewed 805 times
Despite being in the frame under strict AI settings (Night Profile: Min Confidence 60%, Pre-Trigger images: 3, Post-Trigger images: 30, Analyse one image every 100ms), the system failed to detect me in that instance. Strangely, it had successfully detected me when I returned home yesterday evening in the same lighting conditions etc. This inconsistency is bewildering, especially given the meticulously tuned settings aimed at ensuring maximum detection reliability.
The transition from CUDA to DirectML may have influenced the frequency of detections, a shift from the previously uninterrupted detection performance. This change could either be coincidental or a minor glitch that happened to coincide with the move from CUDA to DirectML.
Sharing this concern seemed necessary as it continues to merit further attention.
In my quest for solutions, I stumbled upon a recent topic on here suggesting that "unicorn" is now obsolete. Also, after further digging I found elsewhere that it was suggested adjusting the global AI tab in the camera settings. By entering "Nothing found:0" in the "To cancel" box, it effectively removes "Nothing found" alerts (marked green) from the Confirmed alerts list, compelling the AI to review all images within an alert to identify the best one. Sounds very similar to the Unicorn trick?
It does make me wonder if replacing "unicorn" with "Nothing Found:0" in the cancel field might compel the AI to re-examine the images and successfully identify the object, or is this wrongly assumed ?
I'll conclude my observations here, as I've shared a substantial amount and have other concerns to address in separate posts.
To recap, I am determined to continue utilising the reliability monitor to address any persisting errors and enhance the score as much as possible for system stability purposes and I really need to figure out why CPAI fails to detect sometimes as per the example above.
Regards
SN
