SSL is the synonym for Semi-Supervised Learning. What you refer here is Self-Supervised Learning, which is related to unsupervised learning

Or you could compare your training after e.g. two epochs and only run the best for 500 Epochs

Try to use smaller subsets of your data. It is very likely that the performance scales with the amount of data afterwards.

But if you studied informatics shouldnt you have some sort of coding experiece?

Yes, because Deep Learning is way better than conventional Methods.

No, because the object detector can solve the problem in a single forward path. Todays deep learning based object detectors like Yolo or RCNN + Swin are very good choices for a detection task

Deep Learning Classifiers based on Convolutions also go around the whole image. And the sliding window approaches are not competitive anymore in terms of accuracy as well

Sliding window approches are „Conventional“ Image Processing techniques which are not comptitive anymore nowadays.

I guess yolov7 is a good choice, but depends on your institute be aware of the licenses of the code.

I guess on Coco the best accuracy is given by transformer networks like Swin, but I would assume your dataset is not as big as coco, therefore transformers might not generalize well.

Depends on the dataset. If you would like to detect blue rectangles on a white background you could even achieve 100%.

Than the answer depends on your data I would say

Do you train on COCO?

Maybe with artificial noise? But I am not familiar with how diffusion is really trained, maybe thats how diffusion works.

Would be interesting for a comparison. Since Diffusion Based Models seems like performing bad for faces, it might be interesting to see it perform image to image for synth to real like cycle-gan

I really like the material. I think it would be helpful if you would support solutions for the problems you gibe in each section.