[QUESTION] Did u compare ssl pretrained with coco pretrained for object detection?

Nobody has claimed this yet.

First, I’d like to say thank you — you’re doing a great job, and I truly appreciate the work and effort behind this project.
Question is, did u make comparison with ssl pretrained models and coco pretrained models for object detection?

I have dataset of 30k images of person in hardhat, person without hardhat.
Environment is noisy manufacturing, so it is a big difference from usual coco images in terms of distribution.
First, on unlabeled images trained according to this tutorial trained ssl checkpoint https://docs.lightly.ai/train/stable/models/ultralytics.html

Then trained 2 models, both on 10% of labeled data, one with exported_last.pt from ssl, one with coco pretrained, all other hyperparameters are the same, same full 100% validation.
And i still got better metric with coco, like with 8% F1 difference.
Decreasing amout of labeled data to 2-3-5% didnt help.

Review the comparison described in the issue and the linked Ultralytics SSL training tutorial first. Reproduce the SSL-versus-COCO-pretrained experiment using the stated labeled-data percentages, shared hyperparameters, and full validation set. Done means documenting a supported comparison or a clear explanation of the observed metric difference; no repository files or tests are identified in the issue.

Written by the indexing model from the issue text.