ISIP-CART · HuagoSR · Jul 8, 2026 · Jul 8, 2026 · Jul 8, 2026
diff --git a/android/cartfollow-devlog.md b/android/cartfollow-devlog.md
@@ -762,3 +762,124 @@ $env:Path="$env:JAVA_HOME\bin;$env:Path"
 5. 默认左上角不再显示大块完整 debug；点击 `调试详情` 后可展开完整字段。
 6. Stop / Cancel / Retake / 页面暂停后，诊断 session 应关闭，按钮禁用并复位。
 7. 导出诊断目录后，应能用 `events.csv` 附近的 `frame_log.csv` 和 `identity_log.csv` 判断黄框不转绿或非目标转绿的原因。
+
+---
+
+## 13. Phase C ReID 输入方向修正（2026-07-08）
+
+### 13.1 修正原因
+
+PC 侧 `cartfollow_diagnostics` 复盘发现，旧版诊断 gallery 全部被标记为 `landscape_or_rotated`。代码检查确认旧版 `ReIDCoordinator` 直接在原始 `workingFrame` 上按 bbox 裁剪，并立即送入 `TfliteReIDFeatureExtractor`，没有按 `sensorOrientation` 将 person crop 转正。
+
+这意味着旧版 Android ReID 虽然可以运行并输出分数，但 gallery 与候选 crop 都可能以横向姿态进入模型，影响目标返回和多人干扰场景下的稳定性。
+
+### 13.2 本轮代码变化
+
+- `ReIDCoordinator.collectInitializationCandidate()` 增加 `sensorOrientation` 参数，gallery candidate crop 裁剪后旋转为 upright 再提取 embedding。
+- `ReIDCoordinator.evaluate()` 和内部 ReID candidate 推理同样使用 upright crop。
+- `HumanCartSimulatorFragment` 调用 ReIDCoordinator 时传入当前 `sensorOrientation`。
+- 诊断 `gallery/` 中保存的初始化候选 crop 改为 upright 版本。
+- `session_info.json` 写入 `reid_crop_upright=true` 和 `sensor_orientation`。
+- debug 简洁面板和完整面板增加 `reidCrop=upright`，用于实机确认新版路径已生效。
+
+本轮没有修改 ReID 阈值、belief 阈值、bbox gate、状态机恢复规则或真实底盘控制路径。
+
+### 13.3 构建验证
+
+构建命令：
+
+```powershell
+$env:JAVA_HOME='D:\Java\jdk-17'
+$env:Path="$env:JAVA_HOME\bin;$env:Path"
+.\gradlew.bat :robot:assembleDebug
+```
+
+结果：构建通过。构建日志仍有既有 TensorFlow Lite manifest namespace warning、Kotlin/Javac target warning 和 deprecated Gradle warning，均未阻塞构建。
+
+### 13.4 下一轮手机验收重点
+
+1. Human Cart Simulator debug 应显示 `reidCrop=upright`。
+2. 新采集的 `session_info.json` 应包含 `reid_crop_upright=true`。
+3. 新采集的 `gallery/` 不应再全部被 PC 分析脚本标记为 `landscape_or_rotated`。
+4. 对比旧数据，观察 `target_return` 后 `frames_to_reacquire / frames_to_follow` 是否缩短。
+5. 若 upright crop 后仍大量出现 `belief_high_bbox_failed`，下一轮再处理分状态 bbox gate 和 recoverable stop。
+
+---
+
+## 14. Phase C 新旧诊断数据对比结论（2026-07-08）
+
+### 14.1 分析输入
+
+本轮没有继续改 Android 策略，而是先对新旧 `cartfollow_diagnostics` 做 PC 离线对比：
+
+```text
+old: tools/reid_pc_test/images/cartfollow_diagnostics_old/
+new: tools/reid_pc_test/images/cartfollow_diagnostics/
+```
+
+新版数据来自 ReID crop upright 修正后的 APK，`session_info.json` 中应出现：
+
+```text
+reid_crop_upright=true
+sensor_orientation=90
+```
+
+PC 分析命令：
+
+```powershell
+cd tools/reid_pc_test
+python analyze_cartfollow_diagnostics_v1.py ^
+  --compare-roots old=images/cartfollow_diagnostics_old,new=images/cartfollow_diagnostics ^
+  --output outputs/cartfollow_diagnostics_analysis/compare
+```
+
+### 14.2 对比结果
+
+| 数据 | sessions | target_return | recovered_rate | recovered_fast | recovered_slow | not_recovered | hard_stop | best_mean | margin_mean | bbox_default_rate | gallery_candidate_landscape_rate |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| old | 4 | 11 | 0.5455 | 5 | 1 | 3 | 2 | 0.5017 | 0.3431 | 0.4451 | 1.0000 |
+| new | 2 | 16 | 0.8750 | 11 | 3 | 2 | 0 | 0.5992 | 0.4611 | 0.5485 | 0.0000 |
+
+结论：
+
+- upright crop 修正确实生效：新版 `gallery_candidate_landscape_rate=0.0000`。
+- ReID 分数整体提高：`best_mean` 和 `margin_mean` 都高于旧版。
+- `target_return` 后恢复率提升，且新版暂未出现 `hard_stop_before_return`。
+- 这轮结果说明 ReID 输入方向已不是主要问题。
+
+### 14.3 当前剩余问题
+
+新版主要 blocker：
+
+```text
+candidate_switch_penalty: 15
+belief_high_bbox_failed: 10
+```
+
+含义：
+
+- `candidate_switch_penalty` 说明目标返回或多人干扰时，suspected track 仍容易切换，trackId / lockedTrackId 保护还不够稳。
+- `belief_high_bbox_failed` 说明 ReID / belief 已经有较强证据，但 bbox default/strict gate 不通过，导致黄框迟迟不能转绿。
+
+因此下一轮 Android 工作不应继续优先调 ReID 模型、TFLite 性能或 `bestScore / margin` 阈值，而应聚焦：
+
+1. `TargetTrackManager` 的 track association 稳定性。
+2. locked track 的保留与 suspected track 升级规则。
+3. `FOLLOW` 与 `REACQUIRE/IDENTITY_UNCERTAIN/SEARCH` 使用不同 bbox gate。
+4. recoverable stop 与 hard `STOP` 的边界。
+
+### 14.4 下一轮验收口径
+
+下一轮策略改动后继续采集 `cartfollow_diagnostics`，重点比较：
+
+```text
+recovered_rate
+mean_ms_to_follow
+not_recovered_in_window
+candidate_switch_penalty
+belief_high_bbox_failed
+非目标转绿是否增加
+gallery_candidate_landscape_rate 是否保持 0
+```
+
+如果 `candidate_switch_penalty` 和 `belief_high_bbox_failed` 下降，同时非目标转绿不增加，才说明策略改动真正改善了“目标返回后迟迟不转绿”和“干扰者偶发转绿”两个问题。
diff --git a/android/robot/src/main/java/org/openbot/cartfollow/HumanCartSimulatorFragment.java b/android/robot/src/main/java/org/openbot/cartfollow/HumanCartSimulatorFragment.java
@@ -372,9 +372,10 @@ protected void processFrame(Bitmap bitmap, ImageProxy image) {
             Detector.Recognition largestPerson = selectLargest(mappedRecognitions);
             if (currentState == FollowState.CAPTURE_TARGET) {
               if (reidCoordinator != null) {
-                reidCoordinator.collectInitializationCandidate(workingFrame, largestPerson);
+                reidCoordinator.collectInitializationCandidate(
+                    workingFrame, largestPerson, sensorOrientation);
               }
-              maybeSaveGalleryCandidate(workingFrame, largestPerson);
+              maybeSaveGalleryCandidate(workingFrame, largestPerson, sensorOrientation);
             }
             TargetMatcher.MatchResult legacyMatch =
                 matcher.match(
@@ -389,6 +390,7 @@ protected void processFrame(Bitmap bitmap, ImageProxy image) {
                         currentState,
                         frameW,
                         frameH,
+                        sensorOrientation,
                         legacyMatch.score,
                         legacyMatch.matched,
                         legacyMatch.best);
@@ -616,7 +618,13 @@ private void activateDiagnosticSession() {
     boolean reidAvailable = reidCoordinator != null && reidCoordinator.isAvailable();
     int gallerySize = reidCoordinator == null ? 0 : reidCoordinator.getGallerySize();
     diagnosticSession.writeSessionInfo(
-        diagnosticConfig, detectorName, minConfidence, reidAvailable, gallerySize);
+        diagnosticConfig,
+        detectorName,
+        minConfidence,
+        reidAvailable,
+        gallerySize,
+        true,
+        sensorOrientation);
     resetTargetEventButton();
     Toast.makeText(
             requireContext(),
@@ -703,15 +711,17 @@ private void maybeSaveDiagnostics(
         shouldSaveCrop);
   }
 
-  private void maybeSaveGalleryCandidate(Bitmap frame, Detector.Recognition candidate) {
+  private void maybeSaveGalleryCandidate(
+      Bitmap frame, Detector.Recognition candidate, int sensorOrientation) {
     if (diagnosticSession == null || frame == null || candidate == null) return;
     if (candidate.getLocation() == null) return;
     long now = SystemClock.elapsedRealtime();
     if (lastDiagnosticGalleryMs != 0L
         && now - lastDiagnosticGalleryMs < diagnosticConfig.cropIntervalMs) {
       return;
     }
-    Bitmap crop = cropPerson(frame, candidate.getLocation(), diagnosticConfig.paddingRatio);
+    Bitmap crop =
+        cropPerson(frame, candidate.getLocation(), diagnosticConfig.paddingRatio, sensorOrientation);
     if (crop == null) return;
     lastDiagnosticGalleryMs = now;
     diagnosticSaver.saveGallerySnapshotAsync(
@@ -723,7 +733,8 @@ private static Detector.Recognition recognitionForTrack(TargetTrack track) {
     return track == null || !track.isVisible() ? null : track.recognition;
   }
 
-  private static Bitmap cropPerson(Bitmap frame, RectF bbox, float paddingRatio) {
+  private static Bitmap cropPerson(
+      Bitmap frame, RectF bbox, float paddingRatio, int sensorOrientation) {
     if (frame == null || bbox == null) return null;
     float padX = bbox.width() * paddingRatio;
     float padY = bbox.height() * paddingRatio;
@@ -735,7 +746,15 @@ private static Bitmap cropPerson(Bitmap frame, RectF bbox, float paddingRatio) {
     int height = bottom - top;
     if (width <= 0 || height <= 0) return null;
     try {
-      return Bitmap.createBitmap(frame, left, top, width, height);
+      Bitmap rawCrop = Bitmap.createBitmap(frame, left, top, width, height);
+      int rotation = ((sensorOrientation % 360) + 360) % 360;
+      if (rotation == 0) return rawCrop;
+      Matrix matrix = new Matrix();
+      matrix.postRotate(rotation);
+      Bitmap upright =
+          Bitmap.createBitmap(rawCrop, 0, 0, rawCrop.getWidth(), rawCrop.getHeight(), matrix, true);
+      rawCrop.recycle();
+      return upright;
     } catch (Exception e) {
       return null;
     }
@@ -813,7 +832,7 @@ private void updateDebugInfo(
     String compactInfo =
         String.format(
             Locale.US,
-            "fps=%.1f\nstate=%s\naction=%s\npersons=%d\ntrack=%d locked=%d suspected=%d\nbelief=%.2f\nbest=%.3f margin=%.3f",
+            "fps=%.1f\nstate=%s\naction=%s\npersons=%d\ntrack=%d locked=%d suspected=%d\nbelief=%.2f\nbest=%.3f margin=%.3f\nreidCrop=upright",
             fps,
             state.name(),
             behaviorDecision == null ? "-" : behaviorDecision.selectedAction.name(),
@@ -834,7 +853,7 @@ private void updateDebugInfo(
 
   private String buildIdentityDebugLine(IdentityEvidence identity) {
     if (identity == null) {
-      return "reidAvailable=false\ngallerySize=0\nbestScore=0.000\nsecondScore=0.000\nmargin=0.000\nweak/mid/strong=false/false/false\nbboxDefault=false bboxStrict=false prediction=false\nstableMatchCount=0\ncandidateSwitchCount=0\nreidLatencyMs=0\nreidReason=-\nactiveTrackCount=0\ntrackId=-1 lockedTrackId=-1 suspectedTrackId=-1\ntrackAge=0 missedFrames=0\nbelief=0.00 beliefStable=0 beliefUncertain=0\nbeliefReason=-";
+      return "reidAvailable=false\nreidCrop=upright\ngallerySize=0\nbestScore=0.000\nsecondScore=0.000\nmargin=0.000\nweak/mid/strong=false/false/false\nbboxDefault=false bboxStrict=false prediction=false\nstableMatchCount=0\ncandidateSwitchCount=0\nreidLatencyMs=0\nreidReason=-\nactiveTrackCount=0\ntrackId=-1 lockedTrackId=-1 suspectedTrackId=-1\ntrackAge=0 missedFrames=0\nbelief=0.00 beliefStable=0 beliefUncertain=0\nbeliefReason=-";
     }
     ReIDMatchResult reid = identity.reidMatch;
     BboxContinuityEvidence bbox = identity.bboxContinuity;
@@ -847,7 +866,7 @@ private String buildIdentityDebugLine(IdentityEvidence identity) {
     String reason = reid == null ? identity.reason : reid.reason;
     return String.format(
         Locale.US,
-        "reidAvailable=%s\ngallerySize=%d\nbestScore=%.3f\nsecondScore=%.3f\nmargin=%.3f\nweak/mid/strong=%s/%s/%s\nbboxDefault=%s bboxStrict=%s prediction=%s\nstableMatchCount=%d\ncandidateSwitchCount=%d\nreidLatencyMs=%d\nreidReason=%s\nactiveTrackCount=%d\ntrackId=%d lockedTrackId=%d suspectedTrackId=%d\ntrackAge=%d missedFrames=%d\nbelief=%.2f reidC=%.2f bboxC=%.2f predC=%.2f switchP=%.2f\nbeliefStable=%d beliefUncertain=%d\nbeliefReason=%s",
+        "reidAvailable=%s\nreidCrop=upright\ngallerySize=%d\nbestScore=%.3f\nsecondScore=%.3f\nmargin=%.3f\nweak/mid/strong=%s/%s/%s\nbboxDefault=%s bboxStrict=%s prediction=%s\nstableMatchCount=%d\ncandidateSwitchCount=%d\nreidLatencyMs=%d\nreidReason=%s\nactiveTrackCount=%d\ntrackId=%d lockedTrackId=%d suspectedTrackId=%d\ntrackAge=%d missedFrames=%d\nbelief=%.2f reidC=%.2f bboxC=%.2f predC=%.2f switchP=%.2f\nbeliefStable=%d beliefUncertain=%d\nbeliefReason=%s",
         reidAvailable,
         gallerySize,
         best,

diff --git a/android/robot/src/main/java/org/openbot/cartfollow/ReIDCoordinator.java b/android/robot/src/main/java/org/openbot/cartfollow/ReIDCoordinator.java
@@ -2,6 +2,7 @@
 
 import android.app.Activity;
 import android.graphics.Bitmap;
+import android.graphics.Matrix;
 import android.graphics.RectF;
 import android.os.SystemClock;
 import java.io.IOException;
@@ -77,13 +78,15 @@ public ReIDMatchResult getLastResult() {
     return lastResult;
   }
 
-  public void collectInitializationCandidate(Bitmap frame, Recognition candidate) {
+  public void collectInitializationCandidate(
+      Bitmap frame, Recognition candidate, int sensorOrientation) {
     if (!isAvailable() || frame == null || candidate == null || candidate.getLocation() == null) {
       return;
     }
     if (pendingGallery.size() >= MAX_PENDING_GALLERY) return;
-    Bitmap crop = cropPerson(frame, candidate.getLocation(), 0.08f);
+    Bitmap crop = cropPerson(frame, candidate.getLocation(), 0.08f, sensorOrientation);
     float[] feature = extractor.extract(crop);
+    crop.recycle();
     if (feature != null) pendingGallery.add(feature);
   }
 
@@ -105,11 +108,12 @@ public IdentityEvidence evaluate(
       FollowState state,
       int frameW,
       int frameH,
+      int sensorOrientation,
       float legacyScore,
       boolean legacyMatched,
       Recognition legacyBest) {
     ReIDMatchResult result =
-        maybeRunReID(persons, frame, memory, state, frameW, frameH, legacyBest);
+        maybeRunReID(persons, frame, memory, state, frameW, frameH, sensorOrientation, legacyBest);
     Recognition best = lastBestCandidate != null ? lastBestCandidate : legacyBest;
     boolean matched = legacyMatched;
     float confidence = legacyScore;
@@ -136,6 +140,7 @@ private ReIDMatchResult maybeRunReID(
       FollowState state,
       int frameW,
       int frameH,
+      int sensorOrientation,
       Recognition legacyBest) {
     if (!isAvailable()) {
       lastResult = ReIDMatchResult.unavailable(disabledReason, 0);
@@ -166,8 +171,9 @@ private ReIDMatchResult maybeRunReID(
     long start = SystemClock.elapsedRealtime();
     List<CandidateScore> scores = new ArrayList<>();
     for (CandidateRef ref : candidateRefs(persons, legacyBest)) {
-      Bitmap crop = cropPerson(frame, ref.recognition.getLocation(), 0.08f);
+      Bitmap crop = cropPerson(frame, ref.recognition.getLocation(), 0.08f, sensorOrientation);
       float[] feature = extractor.extract(crop);
+      crop.recycle();
       if (feature == null) continue;
       float score = maxSimilarity(feature, confirmedGallery);
       scores.add(new CandidateScore(ref.index, ref.recognition, score));
@@ -241,7 +247,8 @@ private static List<CandidateRef> candidateRefs(List<Recognition> persons, Recog
     return refs;
   }
 
-  private static Bitmap cropPerson(Bitmap frame, RectF bbox, float paddingRatio) {
+  private static Bitmap cropPerson(
+      Bitmap frame, RectF bbox, float paddingRatio, int sensorOrientation) {
     int fw = frame.getWidth();
     int fh = frame.getHeight();
     float padX = bbox.width() * paddingRatio;
@@ -252,7 +259,15 @@ private static Bitmap cropPerson(Bitmap frame, RectF bbox, float paddingRatio) {
     int bottom = clamp((int) (bbox.bottom + padY), 1, fh);
     int width = Math.max(1, right - left);
     int height = Math.max(1, bottom - top);
-    return Bitmap.createBitmap(frame, left, top, width, height);
+    Bitmap rawCrop = Bitmap.createBitmap(frame, left, top, width, height);
+    int rotation = ((sensorOrientation % 360) + 360) % 360;
+    if (rotation == 0) return rawCrop;
+    Matrix matrix = new Matrix();
+    matrix.postRotate(rotation);
+    Bitmap upright =
+        Bitmap.createBitmap(rawCrop, 0, 0, rawCrop.getWidth(), rawCrop.getHeight(), matrix, true);
+    rawCrop.recycle();
+    return upright;
   }
 
   private static List<float[]> selectDiverse(List<float[]> features, int k) {

diff --git a/...d/robot/src/main/java/org/openbot/cartfollow/diagnostics/CartFollowDiagnosticSession.java b/...d/robot/src/main/java/org/openbot/cartfollow/diagnostics/CartFollowDiagnosticSession.java
@@ -75,7 +75,9 @@ public void writeSessionInfo(
       String detectorModelName,
       float minConfidence,
       boolean reidAvailable,
-      int gallerySize) {
+      int gallerySize,
+      boolean reidCropUpright,
+      int sensorOrientation) {
     JSONObject json = new JSONObject();
     try {
       json.put("session_id", sessionId);
@@ -94,6 +96,8 @@ public void writeSessionInfo(
       json.put("min_confidence", minConfidence);
       json.put("reid_available", reidAvailable);
       json.put("gallery_size", gallerySize);
+      json.put("reid_crop_upright", reidCropUpright);
+      json.put("sensor_orientation", sensorOrientation);
       json.put("device_model", Build.MODEL == null ? "" : Build.MODEL);
       json.put("sdk_int", Build.VERSION.SDK_INT);
     } catch (JSONException e) {