Punch Detection/Classification Data Labeling Instructions

[francislabountyjr]

v1 Dataset — Boxing Punch Annotations 🥊

Here’s how we want the first‑pass dataset to look so we can start training some baseline models.

1. What we’re after

The very first models we’ll train are punch detection and punch classification. Whether that ends up as one multi‑task net or separate heads doesn’t change how we label; the data requirements are the same:

• Detection → we need to know exactly when each punch starts and ends (frame numbers).
• Classification → every punch gets a set of attributes: type of punch, intended target, whether it landed, etc.

2. The rectangles (a.k.a. tracks)

Even though the bare minimum is “start frame + end frame + attributes,” actually tracking the glove across those few frames makes QC and later experiments easier. So:

• Good: drop a tiny box in the corner, toggle it on at the first punch frame, off at the last.
• Better (preferred): keep the box locked on the glove itself for the whole punch. It’s usually just a handful of frames, so this is fast once you get the hang of it.

How to track a punch (quick version)

1. Pick the two‑point rectangle tool in track mode.
2. On the first frame where the glove starts moving, draw the box right over the glove.
3. Tap through the frames, nudging the box to stay on the glove until the punch is done.
4. Fill in the punch attributes while the track is still active.
5. Advance one more frame and hit “switch outside property” (that closes the track).
6. Rinse and repeat for every punch you see.

3. Optional: tracking the fighters themselves

If you feel like going the extra mile, you can throw tracks on the two boxers too. Totally not required right now—there are plenty of off‑the‑shelf person‑tracking models we can bolt on later—but having clean boxer IDs might save us headaches down the road. Feel free to skip for v1 and circle back only if it turns out we need it.

---

That’s it! Keep it simple, stay consistent, and shout if anything’s unclear. Let’s build a killer dataset. 🚀

[production23]

Hi Francis,

Great outline, really clear and practical for building a first-pass dataset. We’ve worked on similar video/action labeling projects, and one thing that helps a lot is having the right tooling to keep everything consistent and efficient from the start.

If you’re open to it, we’d be glad to explore a quick pilot with you to streamline the punch annotation workflow and help get you to a solid baseline dataset faster.

Best, EvoLearns Team

[devinschumacher]

@production23 feel free to connect with and message me via Linkedin