Codeproject AI and false positives

H

heyho

n3wb
Jun 10, 2024
13
5
England
What are everyones experiences with Codeproject and false positive, or more particularly wrongly categorised alerts.

I am currently using YOLOv8, although have also tested with YOLOv5 6.2), and am fairly impressed with its person, dog, cat, car sensing. But i do get a fair few false alerts categorising birds as card and even moths as cars as well. That seems to come mainly from the street camera where there are probably 8-10 stationary cars in view most the time.

I also think YOLOv5 6.2 is slightly better than YOLOv8 as well for correct object detection but it should be the other way around.

My setup consists of:

Server version: 2.6.5
System: Windows
Operating System: Windows (Microsoft Windows 11 version 10.0.22631)
CPUs: AMD Ryzen 5 3600 6-Core Processor (AMD)
1 CPU x 6 cores. 12 logical processors (x64)
GPU (Primary): NVIDIA GeForce GTX 970 (4 GiB) (NVIDIA)
Driver: 560.70, CUDA: 12.6 (up to: 12.6), Compute: 5.2, cuDNN:
System RAM: 16 GiB
Platform: Windows
BuildConfig: Release
Execution Env: Native
Runtime Env: Production
Runtimes installed:
.NET runtime: 7.0.10
.NET SDK: Not found
Default Python: Not found
Go: Not found
NodeJS: Not found
Rust: Not found
Video adapter info:
NVIDIA GeForce GTX 970:
Driver Version 32.0.15.5612
Video Processor NVIDIA GeForce GTX 970
System GPU info:
GPU 3D Usage 1%
GPU RAM Usage 1.1 GiB
Global Environment variables:
CPAI_APPROOTPATH = <root>
CPAI_PORT = 32168
 
Hi, have you thought of training your own model for items unique to your surroundings?

How to train a custom YOLOv5 model using CodeProject.AI Training for YoloV5 6.2 module

Below are some recommendations for training a custom model Stop using all other CodeProject.AI modules (Training a model needs all the resources it can get) Nvidia GPU with as much VRAM is recommended (You can train with a CPU but it will be extremely slow and can take days to have well...
ipcamtalk.com ipcamtalk.com
 
  • Like
Reactions: heyho
cyberwolf_uk said:
Hi, have you thought of training your own model for items unique to your surroundings?

How to train a custom YOLOv5 model using CodeProject.AI Training for YoloV5 6.2 module

Below are some recommendations for training a custom model Stop using all other CodeProject.AI modules (Training a model needs all the resources it can get) Nvidia GPU with as much VRAM is recommended (You can train with a CPU but it will be extremely slow and can take days to have well...
ipcamtalk.com ipcamtalk.com
Click to expand...
Yes I looked at that. I retire in a couple of months so will havce more time to tackle it. Thank you for your response.
 
  • Like
Reactions: cyberwolf_uk
CP.AI seems to think the trees at the sides of my house are people all the time. :) I guess I should train my own model, or just accept it is more creative in seeing faces in the leaves than I am. Ha ha.
 
  • Like
Reactions: heyho
FlyingDog.jpg Capture.JPG
 
  • Like
Reactions: heyho
What I find happens when I analyse an alert is that it highlights a fixed car for example (and shows that outlined in an orange border) but if say a moth flies by (which it doesn't classify) and flies over the sight of the car object then the alert gets triggered and comes back as a car. Is there a way around this? or is that part of training a moth
 
For one, I no longer use the default object models (turn off on the AI settings page). I don't need to be detecting fire hydrants, parking meters, giraffes, frisbees, etc.

Instead I choose one of the custom models (usually ipcam-combined, or ipcam-general) and configure the camera for those specific objects. It not only reduces misidentification, but also improves performance.

But even still, you aren't going to completely eliminate false positives. A cat can look like a person, a bug might be a dog, a shadow might be a person, etc. AI isn't perfect.
 
  • Like
Reactions: heyho
You must log in or register to reply here.
Top Bottom