moku-v4

Go board detector used by Kaya to turn a photo of a goban into a position (SGF). Trained with moku.

It detects three classes: black_stone (0), white_stone (1) and board_corner (2). Kaya computes a homography from the 4 corners and snaps every stone onto the grid.

Files

• model.safetensors, config.json, preprocessor_config.json: 🤗 transformers checkpoint.
• model.onnx: what Kaya runs (ONNX Runtime Web). Input pixel_values (batch, 3, 640, 640), RGB scaled to [0, 1] with no mean/std normalization; outputs logits (batch, 300, 3) (apply a sigmoid) and pred_boxes (batch, 300, 4), normalized (cx, cy, w, h). With the corner head, a third output corner_points (batch, 8, 3) holds the 8 best board-corner peaks (x, y, score), x and y normalized to [0, 1] (class-agnostic). The stone-threshold calibration is baked into logits (offset +0.35 on every class logit), so Kaya's fixed 0.035 threshold needs no per-model setting.

Evaluation

split	mAP@50	stone cdAP	corner R@4	perfect boards	errors / board
validation	0.544	0.773	0.642	42% [31%, 54%]	20.8 [12.8, 29.5]
test	0.585	0.794	0.664	44% [35%, 53%]	18.6 [10.6, 28.6]

Board metrics run Kaya's own post-processing (stone threshold 0.035, corners from head) and compare the resulting position with the one read from the annotations: perfect is the share of boards without a single wrong intersection and no corner more than half a cell off. Intervals are 90% bootstrap CIs over photos.