search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.75k stars 7.96k forks source link

Detect small obj on large images (cartoons) #3062

Open DarLador opened 5 years ago

DarLador commented 5 years ago

Hi,

I'm starting to use your amazing tool (thanks to some great guys here who helped me compile it). Now I'm trying to detect "coin" from large noisy cartoons, like this one: image So I need it to detect the "flying" coins, and maybe some of the coins down there. Some coins are larger then the others, for example: image

The size of my images is varied, and up to 1024 pixels.

I've read some tickets here, like #2399 and #3017, and decided to use yolov3.cfg with: batch=64 subdivisions=64 width=640 height=640 max_batches = 4000 I have many nans during training, the avg is not nan but it's very high: image

Therefore I'm not sure about these properties...

I'm training this dataset on my laptop with NVIDIA GeForce GTX 1070 running Windows 10.

Any advice or guidance would be greatly appreciated!

Thanks!

Christopheraburns commented 5 years ago

How many classes do you have? How many images are you using in your training set? What augmentations have you performed on the images?

AlexeyAB commented 5 years ago

@DarLador Hi,

I have many nans during training, the avg is not nan but it's very high:

This is normal: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

Note: If during training you see nan values for avg (loss) field - then training goes wrong, but if nan is in some other lines - then training goes well.

Just try to train by using simillar images that you want to detect and read these suggestions: https://github.com/AlexeyAB/darknet#how-to-improve-object-detection

DarLador commented 5 years ago

@Christopheraburns currently I have 1 class, but I will have some more. In fact, it there any limitation about the number of classes per model? Is there any minimum or recommended amount of training images? @AlexeyAB At first I tried Yolov2 with a demo data from this tutorial (~ 250 images), and it took 2 hours to reach 2000 iterations. Then, with my real data, I did as you recommended (with Yolov3). I just wanted to test that, so I used only 129 training images, but it took me 16 hours to reach 2900 iteration. The accuracy is good enough for now. I'm very new in this field, and I think maybe I need to find a computer that can do it faster and at an affordable price :)