Why is TensorFlow called “TensorFlow”?

  • The name TensorFlow is derived from the single- and multi-dimensional arrays that neural networks perform operations on, known as “tensors”. Data in a neural network “flows” through the network as its being classified, passing through weighted nodes. Hence TensorFlow. There were apparently multiple projects known as “TensorFlow” that sprung up at the same time.

How many frames of our object is enough to ensure a good model?

  • That’s going to be completely dependent upon the object, poses you’re trying to account for, backgrounds, and lighting conditions. Adding novel objects on top of pre-trained models doesn’t require thousands of training frames, but it does require the RIGHT training frames to create a general “description” of the object. Exceeding 1,000 frames for a single object is likely overkill.

How do I know if my model is trained well?

  • There are a number of metrics that can help you determine when a model begins to converge (where additional training will likely lead to no benefit). Pay special attention to mAP metrics and Loss metrics, you should see those metrics generally settle by around 100 epochs.

Why does my team get a limited amount of model training time?

  • Training in the Google TensorFlow network on GPU resources is not free. Each team is allocated an amount of time based on the costs of using the fixed cloud resources. Our hope is that a team who is congnizant of their training time should be able to get 4-5 models and additional training time on one model with that allocation.

  • It is not possible for a team to “purchase” additional training time for their account. We’re hoping teams will give us feedback on what they feel a reasonable amount of time could be (and let us figure out how to allocate those resources). However, teams who have the capability and resources to clone the open-source fmltc repository that ftc-ml is based on can run their own “instance” of this tool in their own Google Cloud Project. However, swim at your own risk.

Why can’t I seem to get a 100% object detection prediction?

  • Model predictions are never perfect, and attempting to strive for that makes for a really specific and non-generic model. If object detection probability is really high (in the 90-99% range), it might be pointing out that your model may not be as generic as it could be, or is overtrained; it depends on the datasets and what you’re trying to do. Generally after training if your model is predicting all objects above 50% all the time you’re actually doing really well.

I read somewhere about a parameter I can tweak…

  • There are no parameters tweakable in ftc-ml, sorry. It was designed to be simple and easy to use. If you want, feel free to clone your own fmltc repository, modify the code, and deploy it to your own Google Cloud Project instance! However, swim at your own risk.

Can object bounding boxes overlap?

  • Sure, but if you have “blocks” in front of a “ball” such that objects are obscuring each other, just label the parts that are not obscured. Don’t include areas in your bounding box where “there would be the rest of the ball here if it wasn’t obscured by these blocks”

What are the limitations imposed within the ftc-ml tool? (PER TEAM)

  • Max # of Datasets: 20 (you can delete datasets to make more)

  • Max # of Videos: 50 (you can delete videos to upload more)

  • Max # of Videos performing tracking at once: 3 (for multiple logins doing tracking)

  • Max # Bounding Boxes per frame: 10

  • Max Video Limits: 2 Minutes, 1000 frames, 3840 x 2160 resolution, 100MB