feat: apply orientation handling to instance segmentation

Author: NorbertKlockiewicz

apps/computer-vision/components/vision_camera/tasks/InstanceSegmentationTask.tsx

@@ -96,6 +96,7 @@ export default function InstanceSegmentationTask({
   const frameOutput = useFrameOutput({
     pixelFormat: 'rgb',
     dropFramesWhileBusy: true,
+    enablePreviewSizedOutputBuffers: true,
     onFrame: useCallback(
       (frame: Frame) => {
         'worklet';
@@ -105,9 +106,10 @@ export default function InstanceSegmentationTask({
         }
         try {
           if (!instSegRof) return;
+          const isFrontCamera = cameraPositionSync.getDirty() === 'front';
           const iw = frame.width > frame.height ? frame.height : frame.width;
           const ih = frame.width > frame.height ? frame.width : frame.height;
-          const result = instSegRof(frame, {
+          const result = instSegRof(frame, isFrontCamera, {
             confidenceThreshold: 0.5,
             iouThreshold: 0.5,
             maxInstances: 5,
@@ -129,7 +131,13 @@ export default function InstanceSegmentationTask({
           frame.dispose();
         }
       },
-      [instSegRof, frameKillSwitch, updateInstances, activeModel]
+      [
+        instSegRof,
+        frameKillSwitch,
+        updateInstances,
+        activeModel,
+        cameraPositionSync,
+      ]
     ),
   });
 

packages/react-native-executorch/common/rnexecutorch/models/instance_segmentation/BaseInstanceSegmentation.cpp

@@ -6,6 +6,8 @@
 #include <rnexecutorch/ErrorCodes.h>
 #include <rnexecutorch/Log.h>
 #include <rnexecutorch/data_processing/ImageProcessing.h>
+#include <rnexecutorch/utils/FrameProcessor.h>
+#include <rnexecutorch/utils/FrameTransform.h>
 #include <rnexecutorch/utils/computer_vision/Processing.h>
 
 namespace rnexecutorch::models::instance_segmentation {
@@ -112,9 +114,23 @@ std::vector<types::Instance> BaseInstanceSegmentation::generateFromFrame(
     std::vector<int32_t> classIndices, bool returnMaskAtOriginalResolution,
     std::string methodName) {
 
+  auto orient = ::rnexecutorch::utils::readFrameOrientation(runtime, frameData);
   cv::Mat frame = extractFromFrame(runtime, frameData);
-  return runInference(frame, confidenceThreshold, iouThreshold, maxInstances,
-                      classIndices, returnMaskAtOriginalResolution, methodName);
+  cv::Mat rotated = utils::rotateFrameForModel(frame, orient);
+  auto instances =
+      runInference(rotated, confidenceThreshold, iouThreshold, maxInstances,
+                   classIndices, returnMaskAtOriginalResolution, methodName);
+  for (auto &inst : instances) {
+    utils::inverseRotateBbox(inst.bbox, orient, rotated.size());
+    // Inverse-rotate the mask to match the screen orientation
+    cv::Mat maskMat(inst.maskHeight, inst.maskWidth, CV_8UC1, inst.mask->data());
+    cv::Mat invMask = utils::inverseRotateMat(maskMat, orient);
+    inst.mask = std::make_shared<OwningArrayBuffer>(
+        invMask.data, static_cast<size_t>(invMask.total()));
+    inst.maskWidth = invMask.cols;
+    inst.maskHeight = invMask.rows;
+  }
+  return instances;
 }
 
 std::vector<types::Instance> BaseInstanceSegmentation::generateFromPixels(

packages/react-native-executorch/src/modules/computer_vision/InstanceSegmentationModule.ts

@@ -289,6 +289,7 @@ export class InstanceSegmentationModule<
   override get runOnFrame():
     | ((
         frame: Frame,
+        isFrontCamera: boolean,
         options?: InstanceSegmentationOptions<ResolveLabels<T>>
       ) => SegmentedInstance<ResolveLabels<T>>[])
     | null {
@@ -315,6 +316,7 @@ export class InstanceSegmentationModule<
 
     return (
       frame: Frame,
+      isFrontCamera: boolean,
       options?: InstanceSegmentationOptions<ResolveLabels<T>>
     ): SegmentedInstance<ResolveLabels<T>>[] => {
       'worklet';
@@ -340,6 +342,7 @@ export class InstanceSegmentationModule<
 
       const nativeResults = baseRunOnFrame(
         frame,
+        isFrontCamera,
         confidenceThreshold,
         iouThreshold,
         maxInstances,

packages/react-native-executorch/src/types/instanceSegmentation.ts

@@ -211,12 +211,14 @@ export interface InstanceSegmentationType<L extends LabelEnum> {
    * Available after model is loaded (`isReady: true`).
    *
    * @param frame - VisionCamera Frame object
+   * @param isFrontCamera - Whether the front camera is active (for mirroring correction).
    * @param options - Optional configuration for the segmentation process.
    * @returns Array of SegmentedInstance objects representing detected items in the frame.
    */
   runOnFrame:
     | ((
         frame: Frame,
+        isFrontCamera: boolean,
         options?: InstanceSegmentationOptions<L>
       ) => SegmentedInstance<L>[])
     | null;