Return (train, val, test) image counts that sum to n. A subset whose percentage is 0 gets exactly 0 images (this is the fix for the o

Run text-prompted detection. Returns list of dicts: {"class_name": str, "bbox": [x1, y1, x2, y2], "score": float, "label":

Export a SAM fine-tuning dataset: ``images/`` + ``manifest.json``. The manifest is the authoritative training source — per-instance ``bbox``/

Return list of {name, box_thr, txt_thr, nms_thr}.

``{display_name: path}`` for fine-tuned checkpoints, for the SAM selector.

A single foreground point well inside the mask. Uses the distance transform's argmax so the point sits near the medial axis rather than on an

Return the absolute path of the `models/` directory used for ML weights. Resolution strategy (single source of truth used by sam_utils, dino_util

Check if a point is inside a polygon.

Step forward one edit. Returns the snapshot to restore, or None.

Register fine-tuned checkpoints and (re)add them to the SAM dropdown.

Drop any in-progress Detail-% coalescing token (on image switch).

Called when the user selects a different class row.

Free GPU/CPU memory held by the loaded model. Three-step recipe required to actually drop the GPU allocation: 1. Move the underlying

Update the offset for centered image display.

Set up the image label for #40-style handle edits + select the mask.

Return a ready ``SAM2Predictor`` for ``base_model`` (a registry name like ``"SAM 2 tiny"`` or a path to a ``.pt``). Forces predictor

Production import order: torch first, then QApplication. This is what main.py does (see ADR-017).

Clamp a just-edited shape into the image. A move slides the intact shape back inside (size/shape preserving); a resize trims to the border.

Return a flat ``[(name, QImage), …]`` list for batch DINO. Regular images are loaded from disk via PIL → QImage. Multi-dim images con

Wire ImageLabel events to controller slots. ImageLabel does not hold a main_window reference any more — every write path is a Qt signa

``cuda:0 (NVIDIA GeForce RTX 4070)`` — make it obvious the GPU is in use (a bare ``cuda:0`` reads to users like the GPU wasn't detected).

Mark a selected annotation the way the sibling open-garden-planner app does: a dashed selection-blue bounding box plus bright square h

Warn (and let the user back out) when no usable GPU is present.

Load (cache) the Grounding DINO model for ``model_path``.

A per-row 1..100 Detail % spinbox (100 = raw). Disabled for annotations with no polygon to simplify (bbox-only imports).

Switch the UI to a regular image or a slice by name. Returns True if a match was found and the switch was issued. Used by batch-revie

Selection → ready state. Downloads happen lazily on first Detect.

Extract numeric sm values from e.g. ["sm_70", "sm_80", "compute_90"].

Prompt for a new phrase text and replace the row. Mirrors the class-rename flow in annotator_window: validate non-empty, strip, reject

Pick CUDA if available and usable; honour DINO_DEVICE override.

Single DINO inference for one class. Returns (boxes, scores, labels).

Save fine-tuned weights as ``{"model": state_dict}`` and prove they reload through ``SAM(out_path)``. Ultralytics' ``_load_checkpoint

Map every vertex from the old bounding box to the new one (affine scale), so resizing the box scales the whole polygon proportionally.

Preprocess ``image`` and compute encoder features into the predictor. Sets ``pred.batch`` explicitly so ``_prepare_prompts`` maps bbox/point

Flatten ``{class: [ann, ...]}`` into raw instance specs the :class:`SampleGroup` rasterises lazily.

Store results for batch review mode.

Grow the rubber band once the drag clears the click threshold.

Accept current temp_annotations (called from keyPressEvent).