See what the detector actually found, drawn on the footage - #23
Merged
Merged
Conversation
"It hasn't really detected anybody" cannot be answered with track counts, and the question underneath it — is the thing being followed actually my child? — cannot be answered with them either. Both take a second once the boxes are drawn on the video. Each moment's player now carries a canvas overlay fed by a tracks endpoint scoped to that moment's window. Drawn client-side from data rather than burned into a re-encoded video: nothing to render, nothing to wait for, and the focal athlete can be drawn differently from everyone else — bright and thick, while the other players stay deliberately quiet at 45% opacity. Ball, hoop and referee get their own colours and labels. Boxes are stored in source pixels and the player shows a 540p proxy, so everything scales against the frame size the server reports rather than whatever the element happens to be sized at. It also says in words what the boxes show: "12 tracked here, 1 is your athlete, no ball seen" is a diagnosis, where "3755 tracks" was a number. The player itself is no longer folded inside a <details>. That is twice now that something necessary has been hidden behind a summary and gone unnoticed for hours — the identify panel, then this. preload=none already means an unwatched player costs nothing, so hiding it bought nothing and cost an evening. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
You asked whether there is any way to see what the AI detected — boxes, focal points, tracking. Yes, and it answers your other question too.
Boxes on the video
Every moment's player now draws what was tracked at that instant:
Drawn on a canvas from a tracks endpoint scoped to the moment's window, not burned into a re-encoded video. Nothing to render, nothing to wait for, and a Hide boxes toggle.
It also says in words what the boxes show:
That is a diagnosis. "3755 tracks" was a number.
The player is no longer hidden
You said you still could not play a moment without generating clips. The player shipped in #21 — folded inside a collapsed
<details>labelled "Play 2:16 → 2:21", among the Keep/Reject buttons. Easy to miss, and you did.That is twice I have buried something necessary behind a
<summary>: first the identify panel, now this.preload="none"already means an unwatched player costs nothing, so hiding it bought nothing and cost you an evening. It is visible now, and a test asserts it is not inside a<details>.On "maybe we need more footage"
I do not think so, and the boxes will show you why. Your last run:
It is detecting people constantly. What it could not do is tell which person is Sam (fixed), see the ball more than ~7% of the time (improved by tiling), or see the hoop at all (fixed, below). More footage would have produced more of the same.
The basketball model is now live
REELEEL_CV_MODEL=/data/models/bard.onnxis set on the service and confirmed inside the container. Your next detection run uses it — the one whose log you pasted still saidcannot detect referee, hoopbecause it started before this.Measured on your footage on your hardware:
toward_goalandactivity_near_goal— 0.35 of the scoring weight, dark since the first commit — can finally fire.Caveat worth stating: that variable is global, so every project now uses the basketball model regardless of sport. Fine for one project; per-project model selection is the proper fix.
Verification
462 tests pass (2 new), lint and typecheck clean.