stephanecharette
commented
4 years ago TLDR; image [A] is correct. Fix [B] and [C] to be marked up like image [A] and your network will function much better.
These are awesome. Here are some thoughts:
- In image [A] I like how you marked up both the fire and the smoke.
- In image [B] the fire is marked up correctly but I would have done 2 things different with the smoke:
- I would have deleted the smaller smoke region.
- I would have extended the larger smoke region to cover the same area covered by the deleted smoke region, even if it overlaps the fire.
- In image [C] I think you are trying too hard to simply get 100% coverage, which isn't how training works. I can imagine this would simply confuse the neural network.
That last image needs to be explained. When the neural network runs, it needs to detect (for example) where there are flames. In the way you marked it up, there are 6 distinct areas with flames. But the truth is, there is 1 big flame that covers a big part of the image. The neural network should never see that image and think "I need to figure out how to break this out into 6 different fires." That is what you need to mark up, a single flame in picture [C].
Same with the smoke. In that last image, you have 6 smoke areas. I think that is wrong. But with the smoke, it gets subtle and may be difficult. There are a few ways to think about it -- lets go back to image [A].
You want to detect smoke. But do you want to detect all the smoke? Or the difference where the smoke and clear sky meet? I suspect it will be marginally easier to find the edges of the smoke than smoke itself. For example, in image [C], if you look at the branches in the top right of the image...is there smoke? Is it a cloudy day? The top-left definitely has smoke. If I was doing this project and these are the only 3 images I had to determine how I would mark the dataset, I would mark up the "black" smoke of image [C] and not mark up the air around the branches in the top-right corner.
Something important I had written above but cut-and-pasted here instead: [...] hopefully not get confused by cute puffy white clouds or angry storm clouds in the sky. So I hope you also have some images without fires and smoke to train with. This will be critically important, so it doesn't accidentally train on "trees" and start thinking that some trees represent "fire". Here is another example: a beautiful orange sunset over the mountain trees. If you don't include images without fire, you'll end up training on things you don't expect.
B1D1ng
HHHcwb
nanhui69
I am a beginner and am currently using yolov4-tiny for training and want to recognize flames and smoke. If I train 2000 pictures that individually mark the flame, the mAP can reach 75%, but if I train the flame and smoke at the same time, the mAP is less than 50%. The total data set of flames and smoke has 4000 pictures. I think the number of images in the data set is already a lot. Maybe there is a problem with the way I label the images? There are two labeling methods below, which one is appropriate?
A:
B:
Because the range of smoke is relatively large, can it coincide with the flame label when labeling like A? Or is it better to avoid overlapping labels like in B?
Looking forward to your reply, thank you !