| --- |
| license: agpl-3.0 |
| library_name: tf-lite |
| tags: |
| - object-detection |
| - yolov8 |
| - billiards |
| - tflite |
| - cuedetat |
| --- |
| |
| # CueDetat MASTER_POOL_MODEL — Pool Table / Ball Detector |
|
|
| The TFLite detector used by the CueDetat Android app for pocket/table and ball/cue detection. |
| YOLOv8n, 640×640 input, FP16, NMS in-graph. |
|
|
| ## Files |
| - `MASTER_POOL_MODEL.tflite` — a single binary that **concatenates four sub-models**. |
| - `MASTER_POOL_MODEL.tflite.meta` — byte offsets/lengths for each segment (required to load): |
| `MODEL_i_OFFSET` / `MODEL_i_LENGTH`. The app maps each segment to its own interpreter. |
| - `metadata.yaml` — Ultralytics export metadata (classes, imgsz, stride). |
|
|
| ## Segments the app actually runs |
| - Segment 0 — pocket/table detector. Output: `[1, N, 6]` rows `(y1, x1, y2, x2, score, classId)`. |
| - Segment 2 — pool ball / cue detector. Same row layout. |
| - (Segments 1 and 3 exist in the file but are unused by the app.) |
|
|
| ## Classes (from `metadata.yaml`) |
| `0: pool-table 1: pool-table-hole 2: pool-table-side` |
| (plus ball/cue classes used by the pool head: ball, cue). |
|
|
| ## Notes |
| - Input image normalized to [0,1], RGB, 640×640. |
| - License: AGPL-3.0 (Ultralytics-derived). |
|
|
