How to improve the number of stages that a cascade classifier can be trained to?

Python
1

I have a small number of positive samples, only about ten images. But，when I start training, no matter how low I set the number of negative samples, it gets stuck at the step of collecting negative samples after completing 4 rounds of training. What is frustrating is that a cascade classifier trained for only 4 rounds cannot meet my requirements at all. Do you have any good methods to solve this problem, including how to train it for more rounds or other methods to improve the recognition quality of the cascade classifier?

2

how do you expect this to work ?

if you can’t have more (x 1000) positive images, no successful training.

full stop.

what are you even trying to detect ?
in most cases, a cascade classifier isn’t even feasible…

3

I want to detect whether people are wearing school uniforms by detecting the school emblem on the uniforms. I have also tried directly using images of people wearing school uniforms for training, but the results were poor. Currently, I am also trying to use the YOLO model, template matching, and feature matching, but the effectiveness of template matching and feature matching is not good. Training the YOLO model may take some time due to data issues. In the case of a small number of available positive samples, do you have any other recommended methods to achieve this purpose

4

Thanks for your Reply

5

try to get more data (stop fighting that!).

maybe you can finetune yolo with a few hundred positives,
but for cascades, you still need to start from scratch with some thousands.

6

“Thank you for your reply. I have another question: when you mentioned ‘positives,’ are you referring to the training set? What are the recommended ratios for the training set, testing set, and validation set?”

7

yes

for cascades, tons of cropped positives, ad somewhat less (but larger) negatives (it will crop out many tiles for the negs). there’s also no such thing as a test or validation set.

for cnn’s, you will need ‘labelled ground truth data’, the full image (so it contains both fg/bg) + class/bbox information (e.g. in a text file). check out huggingface & roboflow for tools/workflows