I have trained my object detection model on yolov4-tiny and deployed it on oak-d. I am getting 100% detections and no false positives. But when i run the modified gen2-cumulative count script, it misses a lot of the detections and doesnt count them.(50 % not counted but all detected).
I have made the following changes to the oriiginal script:
- Changed input size to 416,416.
- changed the nn pipeline to fit yolov4-tiny.
- put the roi line at 0.9
I can see my detected object passing by with an id but is not added to the cumulative count. This happens mostly when multiple objects pass the roi line at once. There are around 500 objects in the frame(all are detected 100%). but the count only increases to approx 250-300.
And another question: can see the ID on some detections fluctuating meaning it assigned a new id to it? Why does this happen? how to avoid it?
My setup: camera facing down birds view to a steady speed conveyor with black candies on it. The candies are in huge numbers and the yolo-v4 tiny does a very good job at identifying them. I want the total count of the candies that went through the conveyor.
I am using a raspberry pi 4