AlexeyAB
commented
5 years ago @maxwack Hello,
- Usually automatic augmentation is enough. There are no yet only: rotation and vertical flip.
I tried once, to create 4000 images based on the 200 I had by changing exposure,saturation and rotation.But after doing this and training 4000 epochs (2 classes* 2000), the network can not detect anything anymore.
-
If you can't detect anything after 4000 iterations, even on images from training dataset, and even with threshold
-thresh 0.05then the most likely your dataset is broken. Try to check your 4000 images by using Yolo_mark, will be bounded boxes placed correrctly? https://github.com/AlexeyAB/Yolo_mark Also train with flag-mapso you will see accuracy (mAP) during training: https://github.com/AlexeyAB/darknet#when-should-i-stop-training What mAP can you get? -
Looks like your new dataset is broken, try to check it using Yolo_mark.
-
What cfg-file do you use? Can you show screenshot of cloud of points using this command?
./darknet detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 -show
maxwack
Hello @AlexeyAB ,
I have few questions about data augmentation. I read most of the issue on this repository and I know that this project is doing data augmentation based on settings in cfg file.
My problem is that I have a very small data set, around 200 images for 2 classes. The images are medical ultra-sound images so, as you advised in another issue, I set the following in the cfg file : (My images are not X-ray but ultra-sound tho) saturation = 1.05 exposure = 2.5 hue=.01 (https://github.com/AlexeyAB/darknet/issues/1758)
First I want to make clear that I successfully trained the network with this small data set and got decent result (85% of correct classification and localization), but I want to improve this result to 90% or more and was thinking about data augmentation.
1) You advise to have 2000 images for one class. →Do I need to do more data augmentation to reach 2000 images, or automatic augmentation is enough?
2) I tried once, to create 4000 images based on the 200 I had by changing exposure,saturation and rotation.But after doing this and training 4000 epochs (2 classes* 2000), the network can not detect anything anymore. →Is it because I need to train much more because I have much more images? Is it because my data set is getting too different from the images I want to detect which are very close to my original data set? Another reason?
3) I also tried to use 200 images of my data set and 200 unrelated medical images as negative image. For example, I have elbow ultrasound images which I want to detect and I used chest X ray image as negative image. (To do so, I created an empty text file for annotation.) →Same as 2), I got no detection anymore. Any reason you could think of ?
4) Is there anything else I can do to hope having a better result?
Thank you for your help.