Conclusion: You can have lots of fun with plain image data sets, without any labeling at all!

Downloads: PASCAL dataset:

PASCAl is a fun annotated dataset. It contains image data, with object identification and location. For example,

Then I thought, why don’t I have some fun with it?

I have taken the PASCAL dataset and thrown away all annotation. Next I wrote a few custom transformers, so that the pictures will be rotated, and the y values are rotations.

The backbone of the model is pretrained resnet34. Add custom head, find optimum LR, freeze and unfreeze head and fit a few times. After 5 minutes of training, I got some somewhat satisfying results.

Shocking how little time I need to do something cool!

Two things I’d like to point out:

• Any data set will do, really. As long as the pictures are reasonably justified.
• Writing of data loader for pytorch is the thing that took me the most time. The building of model, fine tuning, finding hyper parameters took no time at all.

Similar things you can do at home: Stretch detection! Video processing! Yey! Fun!