The delivery model is now on Github along with exported models for YOLOv5 6.2, YOLOv5 NET, YOLOv8 and YOLOv8 NET
Link: Custom delivery models on VideoDad's Github
People can mask USPS AI detections
- Thread starter rdagger
- Start date
-
- Tags
- blueiris alerts deepstack usps
Link: Custom delivery models on VideoDad's Github
The delivery model is now on Github along with exported models for YOLOv5 6.2, YOLOv5 NET, YOLOv8 and YOLOv8 NET
Link: Custom delivery models on VideoDad's Github
MikeLud1
IPCT Contributor
@ChrisMaunder some new models for you to add to CodeProject.AI install.
Hello VideoDad,
Thank you for your work in producing these models.
I am still running an older version of Blue IRIS and using Deepstack for AI.
I recently installed your USPS.pt model to be able to detect the mail delivery. It runs on a clone that runs only a single delivery model. It worked ok but with lots of false positives. I kept increasing the confidence level until it worked fairly reliable ay 93%. Today it alerted at 93% on my blue recycle bin
The USPS truck often has its door open and that makes it a bit harder to detect.
After reading as many posts as I could find on the subject; I was hopeful that the delivery.pt model had improvements that might help. I have been unable to get the delivery.pt model to work at all on my clone meant for delivery vehicles only.
I had guessed that the YOLOv5 6.2 was what I needed and tried that today without success.
Can you please advise as to which model would be correct for Blue Iris 5.5.9.6 and Deepstack GPU-2021.09.1?
I didn't create the USPS.pt model; I only used many of the images in the training of my delivery models. And yes, I did improve it with more images especially from actual security cameras and with doors open.
I was using the YOLOv5 6.2 model on Deepstack prior to switching to CPAI. I believe it should still work on DS but haven't officially tested it.
First I assume you've got delivery.pt in the correct folder for DS custom models. In my model, all the labels are lowercase so make sure you're looking for "usps" when using "delivery" as the custom model.
If that's all correct and you're still having problems, post your AI settings for the camera. Also, please post the logs for BI when it should be seeing a delivery vehicle.
Thankyou for your reply VideoDad,
My apologies to sstratoti after reviewing that he created the USPS.pt.
Yes I did have the model in the custom models folder ( same one that USPS.pt and all my other previous custom models models are located.)
I did use lower case for the delivery.pt model.
After trying to get it to work without success I found that deepstack stopped working all together. I had to uninstall and re-install to re-establish a working Ai system.
Thankyou for confirmation that the YoloV5 6.2 model is the correct choice. I will try it again.
I had downloaded the delivery.pt model from this site that you had posted and tried that a while back without success. However, DS did not get corrupted during that attempt.
one more question I have:
Now that you have the model on github, is the yolov5 6.2 model the same as the original delivery.pt or different? If they are different I will try both once again.
Thankyou for the work you put into the models. I will give it a more thorough effort now that you have offered your input.
My apologies to sstratoti after reviewing that he created the USPS.pt.
Yes I did have the model in the custom models folder ( same one that USPS.pt and all my other previous custom models models are located.)
I did use lower case for the delivery.pt model.
After trying to get it to work without success I found that deepstack stopped working all together. I had to uninstall and re-install to re-establish a working Ai system.
Thankyou for confirmation that the YoloV5 6.2 model is the correct choice. I will try it again.
I had downloaded the delivery.pt model from this site that you had posted and tried that a while back without success. However, DS did not get corrupted during that attempt.
one more question I have:
Now that you have the model on github, is the yolov5 6.2 model the same as the original delivery.pt or different? If they are different I will try both once again.
Thankyou for the work you put into the models. I will give it a more thorough effort now that you have offered your input.
Just to reiterate, when I'm talking about case, I'm talking about the possible objects (not the filename). In the USPS.pt model, the only object was "USPS". In mine, all the objects are lowercase (e.g. amazon, fedex, usps, etc.).
As for updates, I uploaded each of the versions from before as revisions. So if you look at the history of delivery.pt, you'll see the latest plus the prior versions. Version 4 is the one that was previously posted for download.
I have made a couple updates since then to streamline the model, add more training images, etc.
@MikeLud1, thanks for proposing this to @ChrisMaunder. I'd love them to get included in the CPAI install. Let me know if there is anything you need from me. Also, if I update the GitHub repository in the future, do I need to notify someone or will the install pull the latest each time?
Mike, thanks for previously hosting a copy of delivery.pt and creating a corresponding delivery.onnx for it. I don't know if you want to update them with the latest in your repository now, or just remove them to avoid confusion?
Finally, are there other modules (Coral TPU?) and sizes (medium, large) you are planning to target with CPAI and your models? If so, let me know and I'll try to do the same.
Thanks VideoDad,
I did understand the need for lower case on objects to detect.
I was not familiar enough with Github to understand the release versions. I took a look and saw the original at the bottom and 5 other releases. Please explain what you mean by version 4? I did not see version numbers at the site, but as I said I am not that familiar.
I took the time to do some testing. I was able to get the original at the bottom to work. I was also able to get the 3rd from the bottom to work. (The one where you added the open door images.)
I was unable to get the versions released after that to work. ( higher up on the release page ).
I have no idea whether it is just my install or some thing else. I will stick with the 3rd from the bottom release for now and see how it does going forward.
Just providing some feedback regarding the newer versions.
Thanks again for your help in understanding.
Sorry for being so vague in using the not working comment. I should know better.
I have been using the "Testing and Tuning function- using AI". For the models that work I get a lot of boxes with percentages for what it "sees".
Using the models higher up than the third from the bottom, there are no boxes or percentages displayed. This is using the same alert clips for all testing. I also notice that the computer seems to be struggling or slowed down when running the models that do not present boxes or percentages.
Thanks again, I really appreciate all the work that you and others have put into creating the custom models. This site has been invaluable in helping me get things working.
Keep in mind that the "Analyze with Deepstack/CodeProject" 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 and trigger for a car or person or when it triggered. It should only be used to see what AI can find in that clip, like "hmm I wonder if AI 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 AI and it will show EVERYTHING that AI has in its objects to find that it sees in the clip. This method will show you EVERY ITEM AI searches for.
You need to review the .dat files as that will show you how Deepstack/CodeProject interpreted it.
And as you are seeing, you should really try to limit the number of models a camera will run.
Hello wittaj,
Understood. I have always used the .dat files to review, analyze and improve settings. In this particular case I have no new captures using the new delivery models yet, as I have just installed it.
I was only using the Test and Tune on existing clips to confirm the model was or was not working at all. I will tune it further using new .dat files when new deliveries (.dat files) arrive.
Understood. I have always used the .dat files to review, analyze and improve settings. In this particular case I have no new captures using the new delivery models yet, as I have just installed it.
I was only using the Test and Tune on existing clips to confirm the model was or was not working at all. I will tune it further using new .dat files when new deliveries (.dat files) arrive.
Hi VideoDad I updated the delivery model to the most recent on your Github but I am getting this following error in my CP log.
10:39:37:Object Detection (YOLOv5 6.2): [UnpicklingError] : Unable to load model at C:\Program Files\CodeProject\AI\modules\ObjectDetectionYOLOv5-6.2\custom-models\delivery.pt (invalid load key, '\x0a'.)
10:39:37:Object Detection (YOLOv5 6.2): Unable to create YOLO detector for model delivery
Anything else I can try? Thanks
10:39:37:Object Detection (YOLOv5 6.2): [UnpicklingError] : Unable to load model at C:\Program Files\CodeProject\AI\modules\ObjectDetectionYOLOv5-6.2\custom-models\delivery.pt (invalid load key, '\x0a'.)
10:39:37:Object Detection (YOLOv5 6.2): Unable to create YOLO detector for model delivery
Anything else I can try? Thanks
Can you first verify the checksum on the most recent version of delivery.pt?
Type this command in a command window:
certutil -hashfile delivery.pt SHA1
b99be438e838b7b811f1477a3532a02aa39821b6
That doesn't match with the hash code for the latest version. What's the file size?
You nailed it! My first copy was bad. A fresh delivery.pt did the trick. It is running.
Thank you!