Here are two swings, same hitter, overlaid on each other, one makes contact with a ball in Zone 10, the other, Zone 11 (up+out vs up+in)

Barrels (red/purple) tunnel each other like 2 pitches tunnelling each other. But from overhead view you can see that they break off pretty early in the motion

Exciting new updates in baseball computer vision - I generated this (and many other clips like it) completely automatically through a pipeline I've built

Here's Adrian Del Castillo turning 96 above the zone into a gap double. Auto color-coded too, so the denser centers hold more weight in my model

Putting the pieces together in my computer vision project - I've now taken Bogaerts swing footage, programmed the ability to automatically drill out "donut holes" i.e. the section between XB's arms at contact, then colour code so the centers of the data hold more weight

Finally figured this out on Monday. The below video isn’t perfect (helmet..?), but as long as it works conceptually, I can add more training data + annotations and build it into a smooth pipeline within a week or two

I’ve also never seen a baserunner headfirst slide back to a base that isn’t being thrown to. Watch top left of the screen 🤣

Not sure I’ve ever seen a pitcher shake his head as he comes set. Either you shake off a pitch or you’re pissed about a bad call. But a borderline call went his way the pitch before this

Last pitch he threw today

Is this a “I hate the pitch call but I’ll go along with it anyway?”

made some 3d visualizations of a motion capture session. batter approaching, making, and then following through his contact point

A common issue with using computer vision for biomechanics is that baggy jerseys/pants, shin/elbow guards and long hair become part of the data. To combat this, I've used color gradients on my data - the edges now hold less importance than the denser centers

Here's my FIRST test of my self-trained computer vision segmentation model on completely unseen footage. Overall, SUPER happy with the results! A liiiittle bit of flickering and we lost a lower leg in the follow through. Nothing a bit more training data can't fix. A breakthrough for sure

I'm about 300 frames of manual annotation in to my segmentation project. Should be enough to begin training and automate this (incredibly lengthy) process

Andy Pages - Home Run - 101.6 mph EV - 68.6mph Bat Speed

Now fully annotated via cvat.ai

Made with inspiration from ViTPose in Google Colab. Will put the code on my Github this week

I wonder if Andy Pages knows how much effort I've put into tracing his body to build a training set

Hi Tom!! I'm curious how accurate / reliable the body movements (white dots below) in savant's swing visualizer can be trusted to be? I do notice variation in movement from batter to batter, so clearly there's some specificity to this, but would this be considered accurate for an avg swing?

I can’t sleep so I’m working on computer vision instead

green dots are original marker placement, purple dots are the model's "best guess" at where those body parts should be based on the position of the red dots at that frame. This is a spectacular result

(1/2) How I broke swing biomechanics into entirely data-driven segments, dynamic to each hitter and stance:

- Determine principal features of the swing - the markers that explain the most variance (a lot of biomech data is redundant)
- group the non PFs as isotropic (same in all directions) noise

Reposting a really nice clip from Tanner Stokey / @drivelinebaseball.com that helped me understand the direction of forces in their force plate data