Open reshark opened 5 years ago
@reshark Hi,
Just use it as common 3-channel images.
Can you show examples of your images?
The image is as above. I'd like to hear your opinion.
@reshark It should work perfectly. Just try to wirte your custom data augmentation algorithm (C/Python), which will rotate the images with annotations to achive rotation-invariance.
Hi, @reshark,
If you are worried about another object (e.g. Bottle) being annotated with gun because they overlap, I think this is where the need for varied images come in. You need to have different images of guns that has no bottle isolated with it so that the system will not learn the feature of the bottle.
I am interested on where you got your x-ray dataset. May I know where you got it?
Hi, @kenrubiooo, I am in college now and I am working on a project with the company. I'm looking for ways to improve the detection of metal objects in x-ray equipment.
@AlexeyAB, If an additional channel(gray channel) is applied from 3 channel image, would it be better to use "predefined weights(darknet53.conv.74)"? Or is it better to make a new "weights" from scratch? (The code has been modified to allow for 4 channels. )
Hi, @kenrubiooo, I am in college now and I am working on a project with the company. I'm looking for ways to improve the detection of metal objects in x-ray equipment.
So, the dataset came from the company? Thank you!
@reshark If you changed channels=
in cfg-file,then you should train from scratch.
@AlexeyAB, Thank you!
Hello, AlexeyAB.
I am currently working on a project to apply Yolov3 to X-RAY images. The X-RAY image is 3 channels. I'm trying to find a metal object, but the object I'm looking for on the X-RAY image overlaps with the other object and has a low recognition rate. I'd like to hear your opinion on whether there are any settings that can improve performance.