Client-side stereo detection transfer via ONNX

[mattdawkins]

Warp a detection from one camera to the other and measure its length entirely
client-side (web + Electron renderer), with no backend, by running VIAME's
epipolar template-matching stereo model (stereo measurement "method 1") exported
to ONNX under onnxruntime-web, plus a client-side port of the triangulation.

This is the client counterpart to the desktop backend stereo service: the
desktop ViewerLoader warps and measures via native IPC (stereoTransferLine /
stereoTransferPoints / stereoMeasureLine); this does the equivalent in the
browser so it also works on the web.

What's here (client/dive-common/use/stereo/)

• StereoOnnxMatcher — loads the match ONNX model and warps source points to
  target points (epipolar candidate generation + NCC along the curve).
• calibration.ts / npz.ts — parse .npz/.json calibration in-browser
  (mirrors read_stereo_rig), plus invertRig to swap source/target camera.
• triangulate.ts — two-view triangulation and stereo measurement (length,
  midpoint, range, RMS) + length aggregation, porting
  viame::core::compute_stereo_measurement and aggregate_lengths. No SVD:
  the DLT null vector comes from the adjugate of AᵀA, as in the ONNX graph.
• image.ts — RGBA to BT.601 grayscale (matches OpenCV BGR2GRAY used by the NCC).
• frameSource.ts — read full-res frame pixels from a GeoJS viewer.
• useStereoOnnxTransfer — warp a box, head/tail line or polygon to the other
  camera, measure once both cameras hold a line, and bulk-warp a camera.
• Web glue: platform/web-girder/useStereoOnnxWeb.ts, bound to the Viewer's
  stereo-annotation-complete, stereo-track-linked and the Import menu's
  stereo-warp-imported in the web ViewerLoader.vue.

onnxruntime-web is added as a dependency. The exported model is committed at
client/public/models/stereo_match.onnx, which vite build copies into dist/.
The matcher unit test runs against that served model rather than a second copy.

Matching parameters

Held identical to what the desktop interactive stereo service loads from
configs/pipelines/interactive_stereo_template.conf, so both platforms accept
the same matches. Hidden config there and here — no UI on either side.

Config key	Value	Where it lives here
epipolar_min_disparity / epipolar_max_disparity	2 / 300 px	DEFAULT_RANGE in useStereoOnnxWeb
template_matching_threshold	0.5	DEFAULT_THRESHOLD in StereoOnnxMatcher
uniqueness ratio	0 (disabled)	desktop applies no uniqueness test
template_size	13	baked into the ONNX graph at export
epipolar_num_samples	5000	baked into the ONNX graph at export

The model was re-exported at --template-size 13 for this: the exporter's own
default is 25, so the previously committed model did not match the service.

Behavior parity with desktop

• Boxes and polygons warp once; every warped vertex must be a confident match.
• Head/tail lines re-warp when the source is edited so an auto-generated line
  tracks its source; a human-drawn line is marked stereo_user_line and is
  never overwritten.
• Measurement writes length (and canonical fishLength), midpoint_x/y/z,
  midpoint_range, stereo_rms and track-level avg_length; a length locked
  with length_method = user_set is preserved.
• Progress dialog, error dialog and length snackbar mirror the desktop ones.

Enabling on web

isStereoInteractiveModeEnabled() previously hard-returned
isDesktopRuntime() && …, so stereo-annotation-complete was never emitted off
desktop and this feature was unreachable. Since neither the warp nor the
measurement needs a service any more, that gate is removed and the Stereo
Settings panel is no longer desktop-only.

Testing

• Unit-tested (__tests__/, runs under npm test): calibration parsing,
  grayscale, and point warping validated against the VIAME C++/Python reference
  (~0.25 px) via onnxruntime-web in Node; triangulation and measurement
  validated by projecting known 3D points through the rig, with and without
  distortion, and recovering them (0.01 units at 2.5 m).
• Full suite: 732 tests pass; lint clean; build:web succeeds.

Note on the reference port

plugins/onnx/triangulate.py applies the Lindstrom correction with
p1ᵀ E p2 (left, right), but for E = [T]ₓR the constraint that holds is
p2ᵀ E p1 — verified numerically against the fixture rig. Using the Python's
ordering perturbs exact correspondences and costs ~0.2% in recovered depth. This
port uses the transpose. The VIAME-side script is probably worth the same fix.

Fixed during live testing

• The web glue read viewer.multiCamList?.value / aggregateController?.value,
  but Vue unwraps setup refs on the component instance (the desktop ViewerLoader
  reads them bare). Both were undefined, so every stereo op saw zero cameras
  and returned skipped silently.
• Vite pre-bundled onnxruntime-web into node_modules/.vite/deps, where the
  runtime's sibling .wasm request hit the SPA HTML fallback and failed to
  instantiate — the model never loaded under npm run serve. Excluded from
  optimizeDeps; the production build was already fine.
• A missing calibration or a failed triangulation now surfaces in the error
  dialog instead of no-oping.

Outstanding

• Partly live-tested now. Calibration download, the ViewerLoader bindings
  and client-side measurement are confirmed working in a running web viewer
  against a real stereo dataset (length attributes appear on both cameras).
  The warp path (GeoJS frame-pixel read in
  frameSource.geoViewerToImageElement) is still unverified — it was blocked
  behind the two load bugs fixed in this branch and needs a pass now that the
  runtime loads.
• Segmentation seed transfer is not implemented. Desktop's
  segmentation-stereo-segment warps the seed, runs SAM on the other camera and
  can derive head/tail lines plus the measurement on both. That needs SAM in the
  browser, so type: 'segmentation' returns skipped.
  stereo-annotation-reset and stereo-segmentation-finalize stay unbound on
  web for the same reason — they exist only to undo segmentation side effects.
• Bulk "Warp to All" is slow and image-sequence only. It decodes each frame
  over HTTP because the GeoJS viewer only holds the on-screen frame; video
  datasets have no imageData, so off-screen frames yield nothing.
• Measurement includes distortion; the desktop binding does not.
  compute_stereo_measurement_from_calibration builds
  simple_camera_intrinsics(K) with no distortion coefficients, so on a rig
  with significant distortion the web number will differ from desktop. This port
  follows geometry_numpy.py instead. Reduces to identical results when the
  coefficients are zero.
• The disparity range is scene-dependent and has no UI, matching desktop.
  2–300 px suits the rigs interactive_stereo_template.conf targets, but
  VIAME's batch measurement pipes ship 7–724 for others; a rig outside the
  default silently fails to match until someone passes the range option.
• The exporter's defaults disagree with the service config (--template-size
  25 vs template_size = 13), and export_stereo_mapping.py's self-check
  reports a sub-pixel right_points mismatch at both sizes (0.037 px at 13,
  0.40 px at 25) from epipolar-candidate ties. Pre-existing, and both sizes
  still reproduce the reference correspondence, but the exporter defaults look
  worth reconciling VIAME-side.
• DINO-augmented method 2 exists on the VIAME side and is not wired here.
• No cross-camera undo. Deleting or resetting the source annotation leaves
  the warped feature on the other camera.
• onChange is unused on web — writes rely on the Track notify path to flag
  unsaved changes, same as desktop. Left as an extension point.

[BryonLewis]

I think this is about ready, just want to do a bit more local testing before I merge it in.

[BryonLewis]

While it has some built in testing data, and some complications I think this overal ready based on my testing.