Embedded Files
We used TensorFlow to create an AI that recognizes the three different designs on our custom game sleeve filming video in high and low light with different backgrounds to train the model.
We used TensorFlow to create an AI that recognizes the three different designs on our custom game sleeve filming video in high and low light with different backgrounds to train the model.
First Model
First Model
Our first loss classification graph for our trained AI (PleiadesConeV5) showed that our model wasn't going to be very effective. We reached out to FTC forums to help us problem solve, who gave us feedback and suggestions for improvements. One respondent told us the problem was that we only selected part of the image on the cone rather than the whole object itself.
Final Model
Final Model
Our final loss classification graph showed a successful model (PleiadesConesV13). We also had difficulties labeling the images the same each time and often had to go back through the photos to relabel them. We then uploaded it to the FIRST Machine Learning Toolchain. Using that resource, we were able to integrate our TensorFlow into our autonomous code, which will allow us to park in the correct zone indicated by the cone's rotation at the end of our autonomous program.
Our Custom Cone Sleeve
Our Custom Cone Sleeve
The first design of our cone sleeve we made had images of ice cream cones, octopuses, and stars. Later we read in the FTC Tenserflow documentation that Tensorflow is very good at recognizing visual patterns (such as zebra or giraffe print). With this knowledge, we changed our sleeves to stripes, checkers, and dots.
Page updated
Report abuse