Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# topspin-opentt-ball-subset (private)
|
| 2 |
+
|
| 3 |
+
Derivative training subset for Topspin's M2 ball detector. **Private** —
|
| 4 |
+
not for redistribution.
|
| 5 |
+
|
| 6 |
+
## Source
|
| 7 |
+
|
| 8 |
+
OpenTTGames, linked from <https://lab.osai.ai/>. Each match's source files
|
| 9 |
+
are pulled directly from that site.
|
| 10 |
+
|
| 11 |
+
## Frame-extraction recipe
|
| 12 |
+
|
| 13 |
+
For every frame whose index appears as a key in the match's
|
| 14 |
+
``ball_markup.json`` (i.e. every frame where OpenTTGames has annotated a
|
| 15 |
+
ball centre point):
|
| 16 |
+
|
| 17 |
+
1. Seek the source mp4 to that frame.
|
| 18 |
+
2. JPEG-encode the decoded frame (quality 95).
|
| 19 |
+
3. Emit a YOLO label with a single row: class ``0`` (ball), centre at the
|
| 20 |
+
annotated point, bounding box a fixed 32 px square (16 px half-size)
|
| 21 |
+
normalised against the actual frame dimensions.
|
| 22 |
+
|
| 23 |
+
Frames absent from ``ball_markup.json`` (ball occluded or out of frame)
|
| 24 |
+
are not included in this subset.
|
| 25 |
+
|
| 26 |
+
## Layout
|
| 27 |
+
|
| 28 |
+
```
|
| 29 |
+
images/<match_id>/<frame_index>.jpg
|
| 30 |
+
labels/<match_id>/<frame_index>.txt
|
| 31 |
+
data.yaml
|
| 32 |
+
README.md
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
## Class map
|
| 36 |
+
|
| 37 |
+
| id | name |
|
| 38 |
+
|----|------|
|
| 39 |
+
| 0 | ball |
|
| 40 |
+
|
| 41 |
+
## Generated by
|
| 42 |
+
|
| 43 |
+
``topspin-cv/scripts/prepare_ball_subset.py`` — see the Topspin repo for
|
| 44 |
+
the streaming + idempotent generation pipeline.
|