[PYTHON] real-time-Personal-estimation (What should I do to prevent the estimation of images outside the category) * Failure.

Articles up to the last time

Last time, yolov5 was learned to use GPU using a local machine. https://qiita.com/asmg07/items/0abad3e16886cb60ecef Here, as a test, we trained with epochs 1000 and the result was as follows. As a result of testing based on this model, it turns out that learning is not done well as follows. Images outside the category Images in category The problem that can be seen from this result is that it is necessary to think about what to do in order to prevent the estimation of images outside the category.

Summary of this article

For the time being, I would like to actually try what happens if I increase the number of images to be learned from about 100 to hundreds or thousands.

Annotation

Since the image collection is shown above, I will omit it this time. The article is shown below. Search.py in the article https://qiita.com/asmg07/items/8502fe59b65f92d1e379 Vott is used for annotation. Image data ・ Asuka Saito 350 sheets ・ Yoda Yuki 350 sheets I'm thinking of going. In addition, leave only the images that can be used from the images downloaded by search.py and annotate from there You need to be patient here first because you will do.

About learning

epochs 50 times

epochs 100 times

Conclusion

Nogizaka recognition itself using yolo ended in failure. The reason is that it is excellent in recognizing people, but it is difficult to recognize individuals. I don't know why yet. So I will continue development using yolo, but the Nogizaka recognition project is on hold for the time being. Thank you very much.