Struct MlModel 

pub struct MlModel { /* private fields */ }

Loaded and optimised ONNX model for corner refinement.

The model accepts a batch of f32 intensity patches with shape [N, 1, patch_size, patch_size] (values in [0, 1]) and returns [N, 3] with columns [dx, dy, conf_logit]. Only dx and dy are currently used; conf_logit is ignored.

Implementations

impl MlModel

pub fn load(source: ModelSource) -> Result<Self>

Load and optimise an ONNX model from the given source.

For ModelSource::EmbeddedDefault the embed-model Cargo feature must be enabled; an error is returned otherwise.

Errors

Returns an error if the model file cannot be read, the ONNX graph is malformed, or tract optimisation / compilation fails.

pub fn patch_size(&self) -> usize

Side length (in pixels) of the square intensity patch the model expects.

pub fn infer_batch( &self, patches: &[f32], batch: usize, ) -> Result<Vec<[f32; 3]>>

Run inference on a flat batch of intensity patches.

patches must contain exactly batch * patch_size * patch_size f32 values in [N, 1, H, W] order (values in [0, 1]). Returns one [dx, dy, conf_logit] triple per input patch.

Errors

Returns an error if the slice length does not match batch * patch_size², if the ONNX output shape is unexpected, or if tract inference fails.