| --- |
| license: cc0-1.0 |
| library_name: onnx |
| tags: |
| - billiards |
| - trajectory-prediction |
| - onnx |
| - cuedetat |
| - flow-poke |
| --- |
| |
| # CueDetat MYRIAD — Billiards Trajectory Student |
|
|
| Distilled "Flow-Poke" student model used on-device by the CueDetat Android app to predict a |
| billiard ball's trajectory from a top-down table image and a poke. Archived for preservation; |
| the app has since removed the feature. |
|
|
| ## Files |
| - `myriad_billiard.onnx` — FP16, ~466 KB. Exact file that shipped in the app (`assets/`). |
| - `myriad_student_fp32.onnx` (+ `myriad_student_fp32.onnx.data`) — FP32 source in ONNX |
| external-data format; keep both files together. |
|
|
| ## Inputs / Outputs |
| - `image`: Float32[1,3,128,128], CHW, normalized to [-1, 1]. |
| - `poke`: Float32[1,4] = (x, y, dx, dy); position normalized [0,1], displacement ≈ 0.02–0.075. |
| - `trajectory` (output): Float32[1,30,2] = 30 (x,y) points of the poked (ball-0) ball, normalized [0,1]. |
|
|
| ## Expected input frame (top-down synthetic) |
| Grey surround RGB(128,128,128); white table rectangle RGB(255,255,255) inset 6–12 px; balls as |
| solid circles, radius ≈ 4 px (0.0333·128); the **poked ball red** (255,0,0), all others black (0,0,0). |
|
|
| ## Provenance |
| - Teacher: [CompVis/flow-poke-transformer](https://github.com/CompVis/flow-poke-transformer) (billiards physics). |
| - Distillation dataset: Kaggle `hereliesaz/cuedetat-trajectories` (~50k image/poke/trajectory tuples). |
| - Architecture: CNN(image)→128 ⊕ MLP(poke)→64 → MLP[256,128] → [30,2]. |
| - License: CC0-1.0. |
|
|
