The confusion caused by adding background data to the labeled training set.

[Minxiangliu]

Hi @glenn-jocher ,

I've tried your suggestion of labeling the background data and adding it to the training set. However, I encountered some confusion:

1. I've confirmed that all data have corresponding label files (txt files), but the log always shows 1 background. Is this a preset?

train: Scanning /DataSets/labels... 1937 images, 1 backgrounds, 0 corrupt: 100%|██████████| 1937/1937 [00:01<00:00, 1431.80it/s]
train: New cache created: /DataSets/labels.cache
train: Caching images (4.4GB True): 100%|██████████| 1937/1937 [00:02<00:00, 851.14it/s]
val: Scanning /DataSets/labels... 484 images, 0 backgrounds, 0 corrupt: 100%|██████████| 484/484 [00:00<00:00, 782.73it/s]
val: New cache created: /DataSets/labels.cache
val: Caching images (1.1GB True): 100%|██████████| 484/484 [00:01<00:00, 284.94it/s]

2. The confusion matrix outputted by the model includes an additional background category. Is there a way to merge this?
   

3. The model makes two judgments regarding background data—one is with no labels, and the other labels the entire image as the background. Is this a normal result? Is there a way to improve this, or could you offer some advice?
   

I look forward to your response and sincerely appreciate your help in advance!