Looking for an Technical Design Recommendation (e.g., when to use Workflows)

[pnewell4]

I had originally planned/hoped to build the following idea entirely client-side, so I’m curious how experienced developers would recommend designing it:

As a lacrosse coach, we preach that players play “Wall Ball”, or throw/catch a ball off a wall or bounce-back. I want to gamify the experience by counting the number of bounces in a set amount of time. Ideally, this application would have real-time feedback (imagine a sound playing when you break a personal record), but can live with it working in the cloud and processing the video once time has expired.
I tried several, less CV approaches, like diffing frames to detect motion, but I had difficulty filtering out the noise, so I decided to train a custom model. In the videos, the ball is typically a blur across the screen… for starters, does this raise any red flags? Next, tracking will be important to detect whether it bounces and therefore counts toward your score.
I just watched the demo of the “Split the G” application and liked how there is some client-side work and some server-side work. Could that work here?

reading this back, I don’t mean to ask that someone do all the work in designing this thing, I’m more looking for suggestions/nudges in the right direction.

thank you!!

• Project Type: - Object detection w/tracking
• Operating System & Browser: - Browser-based application
• Project Universe Link or Workspace/Project ID: Sign in to Roboflow

[Jacob_Witt]

Hey @pnewell4!

Couple of thoughts:

1. Good rule of thumb is “if a human can see it, the model can see it”. I would just make sure to label each image exactly how you’d like the model to return it to you.
2. The good news on blur is that you have a lot of options for tracking. You could track when the player bounces the ball, or look for a pattern of “catch” and “throw” stances from players (this is probably easier and you could do it with basic object detection).
3. I always recommend starting with the easiest possible version of the solution, so in your case it would be local video capture + run the model in the cloud. Once you have something that works, I’d then spend the time getting it to run locally.