search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.63k stars 7.95k forks source link

How to reduce false detections? #7884

Open scianand opened 3 years ago

scianand commented 3 years ago

Hi @AlexeyAB @WongKinYiu,

I am training a drone detection yolov4 model. I have around 80000 images in total of which 12000 images are negative images of birds. Can you please suggest to me how many iterations I should train it for? I have trained some models and these are still picking up some birds as drones and other than that some flies are also got detected as drones. I have only one class - drone. Please can you suggest to me a way to reduce false positives? I am training the model with 1184x640 network size as in some images drones are very small.

Thank you in advance.

stephanecharette commented 3 years ago

You don't say what size are the images and the drone, other than "in some images drones are very small".

Have you considered using image tiling? E.g., https://www.ccoderun.ca/darkhelp/api/Tiling.html and https://www.ccoderun.ca/darkmark/ImageSize.html ?

scianand commented 3 years ago

The image size is 1920x1080 and the size of the drone is 25x25.

scianand commented 3 years ago

The image size is 1920x1080 and the size of the drone is 25x25.

stephanecharette commented 3 years ago

Image = 1920x1080 Network = 1184x640 Object = 25x25

Once resized, your drones are 15x15. This should definitely be do-able.

What config are you using? What max batches are you using if you have 80K images?

scianand commented 3 years ago

max_batch is 65000 for now.