chuanenlin/drone-net

Wrong detection results

iamrajee opened this issue · 0 comments

Hi, I'm using default cfg and weight for drone detection. However, I'm not getting accurate results. In fact, there are multiple drones are getting detected in a single yolo bounding box.

Screenshot from 2021-11-01 15-01-50
Screenshot from 2021-11-01 15-01-47
Screenshot from 2021-10-31 11-34-23

Here are some of the result on the training images:
predictions
predictions
predictions
predictions

I even trained the network from scratch(MSE Loss ~ 0.18) but still no improvement. I'm using the Yolov4 branch from AlexeyAB/darknet, which expects a normalised coordinate and I'm also using normalised coordinate for training. What would be the reason for my error? Thanks!