Tipps for getting started with object/keypoint detection

im planning to take on an object/keypoint detection problem and was wondering if someone can recommend a good (code-driven) tutorial for single shot detectors like yolo or ssd.

I will build my own dataset and thought about applying transfer learning. Although I would also like to try to train a new model from scratch.
I’m only starting to study object detection methods. However, I find it a bit challenging to understand the training step (format of labels, network output, loss function, etc.).

I thought maybe someone in this course knows a few links that could help someone getting started on object detection :grinning:

Jeremy covered this in his 2018 course that I just stumbled on earlier this week. It needs some adaptation to the current version of the fast ai library. But he gives a beautiful and incremental way to understand how to go from object classification to object detection (single and multiple… Yolo and ssd)

An important revelation to me was how important the loss function is (as compared to the architecture of the model) to tackle new types of tasks.

Two classes starting here.

I found the GitHub repo with the notebooks… Don’t have it handy right now so let me know if you have trouble finding it.

1 Like

Not sure if this is what you are looking for, but I have an object detection tutorial that uses the Icevision library

1 Like

This one is great:

Object Detection from scratch - Single Shot Detector

by @dhoa

Object Detection from scratch - Single Shot Detector


Thank you for the replies :slight_smile:

@Chris-hughes has written a nice deep dive on Yolo7, it is worth a read to get a deeper understanding of all its parts.
YOLOv7: A Deep Dive into the Current State-of-the-Art for Object Detection


As I mentioned earlier, Jeremy gives an awesome introduction to Object Detection in his 2018 course that you can find on youtube. I really appreciate the incremental way that he builds up an intuitive understanding. The only downside is it’s using an older version of the fast ai library.

So I started a new blog series using Jeremy’s stepwise approach but updating everything to use the current version of FastAI. Where I incrementally show you how to build a YOLO/SSD single-shot detector. It will probably be 4 or so articles…

  1. Get a suitable dataset (Pascal VOC) and build a largest object classifier
  2. Build a bounding box predictor for the largest object.
  3. Combine largest object classification and bounding box prediction.
  4. Full Single Shot Detector (SSD/YOLO)

I’ve tweeted it out here if anyone you would like to check it out and maybe give me a like :slight_smile:


Part 2 of my Object Detection Series outline above… if you’d like to give me a like :slight_smile: