Skip to content

Latest commit

 

History

History
9 lines (6 loc) · 1.54 KB

README.md

File metadata and controls

9 lines (6 loc) · 1.54 KB

YOLOv8 Models for IP Camera AI Detection

The easy way to get started with object detection and lots of other self-hosted AI projects is using CodeProject.AI Server.

ipcamv8s-pva - Detects persons, vehicles, and animals. This model was trained using yolov8s (small) parameters to make it quick. The COCO 2017 dataset was refined to include backgrounds and images with persons, vehicles, and animals. Any image which contained one of these objects that was less than 0.05% or ~200px in area (i.e., 20px x 10px) were removed from the dataset. For reference this removed 6,761 person objects however there were still over 224,978 person objects remaining in the COCO Training Dataset. I also used around 500 images from my own camera images which were labeled with Label Studio. The same process was also used for the validation dataset.

ipcamv8n-pva - Detects persons, vehicles, and animals. Nano model. This model was trained on the same dataset as the small version with a patience of 20 for 300 full epochs. I resumed training after the 300 however no improvements were seen. I have tested for a day primarily with Blue Iris and am pretty happy with the results for a model of this size. It does seem to fail to recognize distant animals at times. I think one issue with this may be the large variety of animals. If there is interest I will consider retraining without the exotic animals that the COCO dataset includes.