I built my first set of models. I want to recognize soda cans and coffee cups in a cluttered inside environment. I made about 300 images with the camera on my robot and used them in various permutations to train a yolo9 model. You can see the “results” below. I dont know how to evaluate that. The recognition is ok but I don’t know if the training was right. For example my testing so far is within a typical environment that I want but it’s not the only one. (E.g. Living room vs. bedroom.) Any tips for training and evaluating training?
Hey, that’s a great start! Here’s a few ideas to keep you moving on things:
-
Click on the latest model and review the info provided by Roboflow. There’s lots of good information like “Model Improvement Recommendations” shown below. There are ideas about things to try.
-
The confusion matrix is really handy when you just do not know what to do to improve things. For example, I clicked on “False Negatives” for sheep and it pulls up images where the model missed finding sheep. Sometimes you see unique images and realize you need more examples of that to train on. Or you see they are tiny meaning you might need SAHI or similar to find those scenarios.
-
You mention it’s not the only environment. It’s always best to include examples from every scenario that will be encountered if you can. But there are two other strategies besides brute-forcing it like that. First you can just run your model on new scenarios and see if it works! Sometimes the model has generalized well enough it will work across different environments. Second, you can improve the model. Either save images that “fail” when you run the model. Or “turn on” active learning - this is where you put a “Roboflow Dataset Upload” block into a workflow that is running your model on images and it will send images to your dataset so they can be used in the next round of training.
Hope some of that helps!
