Move the panorama keypoint detector onto the engine, and probe with a detector

XFeat's two exports are a Keypoints role now; the crate no longer names
tract, and the app compiles TensorRT engines for both ahead of the
first merge. The probe picks the smallest *detector* rather than the
smallest file: the tablet's first run chose the 112 KB eye classifier,
which has no int8 form, and reported the Hexagon as failed for want of
one.

core/dr-pano/Cargo.toml

@@ -12,11 +12,12 @@ thiserror.workspace = true
 log.workspace = true
 
 # Inference for the learned keypoint detector, on the same footing as
-# `dr-segment`: `ort` is the API, tract is the engine, and both are optional
-# so that the geometry — matching, the rotation solve, the projections — is a
-# dependency-free crate that tests without a model.
+# `dr-segment`: `ort` is the API, `dr-inference-engine` decides what runs
+# it (docs/inference.md), and both are optional so that the geometry —
+# matching, the rotation solve, the projections — is a dependency-free crate
+# that tests without a model.
 ort = { workspace = true, optional = true }
-ort-tract = { workspace = true, optional = true }
+dr-inference-engine = { workspace = true, optional = true }
 ndarray = { workspace = true, optional = true }
 
 [dev-dependencies]
@@ -30,7 +31,7 @@ default = ["xfeat", "embedded-model"]
 
 # The XFeat detector (FR-MRG-8). Off, the crate has no model and no runtime,
 # and `Detector` has no implementation — a build that only wants the geometry.
-xfeat = ["dep:ort", "dep:ort-tract", "dep:ndarray"]
+xfeat = ["dep:ort", "dep:dr-inference-engine", "dep:ndarray"]
 
 # Compile the weights into the binary, for the same reason `dr-segment` does:
 # Android hands the app no path to read a model from (ARCH §6.9).

core/dr-pano/src/lib.rs

@@ -63,3 +63,13 @@ pub enum PanoError {
     #[error("inference: {0}")]
     Inference(#[source] ort::Error),
 }
+
+#[cfg(feature = "xfeat")]
+impl From<dr_inference_engine::Error> for PanoError {
+    fn from(e: dr_inference_engine::Error) -> Self {
+        match e {
+            dr_inference_engine::Error::Inference(e) => PanoError::Inference(e),
+            dr_inference_engine::Error::Io(e) => PanoError::ModelRead(e),
+        }
+    }
+}

core/dr-pano/src/xfeat.rs

@@ -2,10 +2,11 @@
 //! The XFeat detector — the network under tract, and the decoder after it.
 //!
 //! Apache-2.0 weights (`models/LICENCE.md`), exported at a fixed shape by
-//! `tools/export-xfeat.sh` and loaded through the same `ort`-over-tract
-//! backend `dr-segment` and `dr-face` use, so this adds no runtime and no C
-//! to the tree. ~300 ms per frame on the reference desktop, ~400 ms on the
-//! tablet (S15.2, S15.4).
+//! `tools/export-xfeat.sh` and loaded through the same `dr-inference-engine`
+//! `dr-segment` and `dr-face` use, so this adds no runtime and no C to the
+//! tree; what runs it is the device's business (docs/inference.md). ~300 ms
+//! per frame on tract on the reference desktop, ~400 ms on the tablet
+//! (S15.2, S15.4).
 
 use crate::features::{decode_xfeat, DecodeOptions, Features, XFeatMaps, DESCRIPTOR_LEN};
 use crate::image::Gray;
@@ -30,11 +31,18 @@ const EMBEDDED_PORTRAIT: &[u8] = include_bytes!("../../../models/keypoints/xfeat
 
 /// A loaded detector: the network at both shapes.
 pub struct XFeat {
-    landscape: ort::session::Session,
-    portrait: ort::session::Session,
+    landscape: dr_inference_engine::Model,
+    portrait: dr_inference_engine::Model,
     pub options: DecodeOptions,
 }
 
+/// The bytes of both exports compiled into the binary, for whoever compiles
+/// engines ahead of the first request (docs/inference.md §6).
+#[cfg(feature = "embedded-model")]
+pub fn embedded_model_bytes() -> [&'static [u8]; 2] {
+    [EMBEDDED_LANDSCAPE, EMBEDDED_PORTRAIT]
+}
+
 impl XFeat {
     /// The weights compiled into the binary.
     #[cfg(feature = "embedded-model")]
@@ -53,16 +61,10 @@ impl XFeat {
     }
 
     pub fn from_bytes(landscape: &[u8], portrait: &[u8]) -> Result<Self, PanoError> {
-        install_backend();
-        let session = |bytes: &[u8]| {
-            ort::session::Session::builder()
-                .map_err(PanoError::Inference)?
-                .commit_from_memory(bytes)
-                .map_err(PanoError::Inference)
-        };
+        use dr_inference_engine::{Form, Role};
         Ok(XFeat {
-            landscape: session(landscape)?,
-            portrait: session(portrait)?,
+            landscape: dr_inference_engine::open(Role::Keypoints, Form::F32, landscape)?,
+            portrait: dr_inference_engine::open(Role::Keypoints, Form::F32, portrait)?,
             options: DecodeOptions::default(),
         })
     }
@@ -75,11 +77,13 @@ impl XFeat {
     /// so a caller that already scaled a frame to a proxy maps them on with
     /// the scale it used and nothing else.
     pub fn detect(&mut self, image: &Gray) -> Result<Features, PanoError> {
-        let ((in_w, in_h), session) = if image.height > image.width {
-            (INPUT_PORTRAIT, &mut self.portrait)
+        let ((in_w, in_h), model) = if image.height > image.width {
+            (INPUT_PORTRAIT, &self.portrait)
         } else {
-            (INPUT_LANDSCAPE, &mut self.landscape)
+            (INPUT_LANDSCAPE, &self.landscape)
         };
+        let acquired = model.acquire()?;
+        let mut session = acquired.lock();
         let (fitted, scale) = image.fitted(in_w, in_h);
         let padded = fitted.padded(in_w, in_h);
 
@@ -145,13 +149,3 @@ impl XFeat {
         Ok(features)
     }
 }
-
-fn install_backend() {
-    use std::sync::Once;
-    static ONCE: Once = Once::new();
-    ONCE.call_once(|| {
-        // False if another crate installed it first, which is fine: there is
-        // one backend compiled in for it to have chosen.
-        let _ = ort::set_api(ort_tract::api());
-    });
-}