feat(perception): add DDRNet full-video vegetation replay

2026-08-28 01:13:18 +03:00
parent 67e6ae98e2
commit 30080c51aa
10 changed files with 729 additions and 64 deletions
@@ -97,7 +97,16 @@ function SpatialState({ message: text }: { message: string }) {

 export interface M4ReplayThreatSemanticLayer {
  resultId: string;
-  taxonomy: readonly E47SemanticClass[];
+  spatialResultId?: string | null;
+  maskUrl?: (sequence: number) => string;
+  label?: string;
+  maskAriaLabel?: string;
+  taxonomy: readonly {
+    classId: number;
+    label: string;
+    disposition: "labeled" | "ambiguous" | "prediction" | "undefined";
+    colorRgb: readonly [number, number, number];
+  }[];
 }

 export interface M4ReplayThreatReviewAnchor {
@@ -153,6 +162,8 @@ export function M4ReplayThreatVisual({
  evidenceLabel = "M4.6",
  initialSpatialMode = null,
  classifiedSpatialLayer,
+  showReferenceMediaLayers = true,
+  showSpatialOverlaySummary = true,
  onActiveSequenceChange,
 }: {
  resultId: string;
@@ -164,6 +175,8 @@ export function M4ReplayThreatVisual({
  evidenceLabel?: string;
  initialSpatialMode?: LaboratoryMetricSceneMode | null;
  classifiedSpatialLayer?: M4ReplayClassifiedSpatialLayer;
+  showReferenceMediaLayers?: boolean;
+  showSpatialOverlaySummary?: boolean;
  onActiveSequenceChange?: (sequence: number | null) => void;
 }) {
  const [mediaMode, setMediaMode] = useState<M4ThreatMediaMode | null>("video");
@@ -280,16 +293,30 @@ export function M4ReplayThreatVisual({
      ? lastSpatialFrameRef.current.frame
      : null;
  const cameraPointOverlay = useM4ThreatCameraPointOverlay({
-    enabled: showMediaPoints,
+    enabled: showReferenceMediaLayers && showMediaPoints,
    resultId,
    sequence: frame?.sequence ?? null,
    endpointRoot: timelineEndpointRoot,
  });
+  const semanticSpatialResultId = semantic
+    ? semantic.spatialResultId === undefined ? semantic.resultId : semantic.spatialResultId
+    : null;
+  const spatialSemanticTaxonomy = useMemo<readonly E47SemanticClass[]>(
+    () => semanticSpatialResultId && semantic
+      ? semantic.taxonomy.map((item) => ({
+          classId: item.classId,
+          label: item.label,
+          disposition: item.disposition === "ambiguous" ? "ambiguous" : "labeled",
+          colorRgb: item.colorRgb,
+        }))
+      : [],
+    [semantic, semanticSpatialResultId],
+  );
  const semanticTimeline = useE47SemanticTimelineFrame({
-    resultId: semantic?.resultId ?? null,
+    resultId: semanticSpatialResultId,
    activeSequence: frame?.sequence ?? timelineFrame.activeSequence,
    frameCount: metadata.timeline?.frameCount ?? 0,
-    taxonomy: semantic?.taxonomy ?? [],
+    taxonomy: spatialSemanticTaxonomy,
  });
  const displayingBufferedFrame = Boolean(
    frame
@@ -339,7 +366,8 @@ export function M4ReplayThreatVisual({
  const staticObstacleBoxes = useMemo<readonly RecordedEvidenceBox[]>(() => {
    const timeline = metadata.timeline;
    if (
-      !frame
+      !showReferenceMediaLayers
+      || !frame
      || !timeline?.cameraObstacleProjectionDelivery
      || !showStaticObstacles
    ) return [];
@@ -348,14 +376,14 @@ export function M4ReplayThreatVisual({
      timeline.imageWidth,
      timeline.imageHeight,
    );
-  }, [frame, metadata.timeline, showStaticObstacles]);
+  }, [frame, metadata.timeline, showReferenceMediaLayers, showStaticObstacles]);
  const activeBoxes = useMemo(
-    () => classifiedSpatialLayer ? [] : [
+    () => classifiedSpatialLayer || !showReferenceMediaLayers ? [] : [
      ...boxes(frame?.cameraProposals ?? []),
      ...staticObstacleBoxes,
      ...reviewAnchorBoxes,
    ],
-    [classifiedSpatialLayer, frame, reviewAnchorBoxes, staticObstacleBoxes],
+    [classifiedSpatialLayer, frame, reviewAnchorBoxes, showReferenceMediaLayers, staticObstacleBoxes],
  );
  const semanticClasses = useMemo<readonly RecordedEvidenceSemanticClass[]>(
    () => semantic?.taxonomy.map((item) => ({
@@ -367,10 +395,14 @@ export function M4ReplayThreatVisual({
  const semanticPalette = useMemo<readonly RecordedEvidenceSemanticPaletteEntry[]>(
    () => semantic?.taxonomy.map((item) => ({
      classId: item.classId,
-      color: item.disposition === "ambiguous"
+      color: item.disposition === "undefined"
+        ? { kind: "transparent" as const }
+        : item.disposition === "ambiguous"
        ? { kind: "token" as const, token: "--nodedc-warning-rgb" as const }
        : { kind: "diagnostic" as const, rgb: item.colorRgb },
-      opacity: item.disposition === "ambiguous" ? 0.52 : 0.92,
+      opacity: item.disposition === "undefined"
+        ? 0
+        : item.disposition === "ambiguous" ? 0.52 : 0.92,
    })) ?? [],
    [semantic?.taxonomy],
  );
@@ -580,15 +612,17 @@ export function M4ReplayThreatVisual({
  const semanticOverlay: RecordedEvidenceSemanticOverlay | undefined =
    semantic && showMediaSemantic && frame
      ? {
-          src: e47SemanticMaskUrl(semantic.resultId, frame.sequence),
+          src: semantic.maskUrl?.(frame.sequence)
+            ?? e47SemanticMaskUrl(semantic.resultId, frame.sequence),
          prefetchSrcs: Array.from({ length: 12 }, (_, index) => index + 1)
            .map((offset) => frame.sequence + offset)
            .filter((sequence) => sequence < (metadata.timeline?.frameCount ?? 0))
-            .map((sequence) => e47SemanticMaskUrl(semantic.resultId, sequence)),
+            .map((sequence) => semantic.maskUrl?.(sequence)
+              ?? e47SemanticMaskUrl(semantic.resultId, sequence)),
          classes: semanticClasses,
          palette: semanticPalette,
          opacity: 0.9,
-          ariaLabel: `E47 semantic mask frame ${frame.sequence + 1}`,
+          ariaLabel: `${semantic.maskAriaLabel ?? "Semantic prediction"} frame ${frame.sequence + 1}`,
        }
      : undefined;
  const accumulatedCameraPoints = cameraPointOverlay.overlay?.sequence === frame?.sequence
@@ -659,8 +693,8 @@ export function M4ReplayThreatVisual({
  );

  const mediaLayerControls = semantic
-    || metadata.timeline?.cameraPointDelivery
-    || metadata.timeline?.cameraObstacleProjectionDelivery ? (
+    || (showReferenceMediaLayers && metadata.timeline?.cameraPointDelivery)
+    || (showReferenceMediaLayers && metadata.timeline?.cameraObstacleProjectionDelivery) ? (
    <div
      className="m4-replay-threat-visual__pane-layer-controls"
      role="group"
@@ -677,7 +711,7 @@ export function M4ReplayThreatVisual({
          SEMANTICS
        </Button>
      ) : null}
-      {metadata.timeline?.cameraPointDelivery ? (
+      {showReferenceMediaLayers && metadata.timeline?.cameraPointDelivery ? (
        <Button
          size="compact"
          shape="pill"
@@ -689,7 +723,7 @@ export function M4ReplayThreatVisual({
          POINTS
        </Button>
      ) : null}
-      {metadata.timeline?.cameraObstacleProjectionDelivery ? (
+      {showReferenceMediaLayers && metadata.timeline?.cameraObstacleProjectionDelivery ? (
        <Button
          size="compact"
          shape="pill"
@@ -738,7 +772,7 @@ export function M4ReplayThreatVisual({
      >
        {classifiedSpatialLayer.cellLayerLabel}
      </Button>
-      {semantic ? (
+      {semanticSpatialResultId ? (
        <Button
          size="compact"
          shape="pill"
@@ -796,7 +830,7 @@ export function M4ReplayThreatVisual({
          LOW-STEP
        </Button>
      ) : null}
-      {semantic ? (
+      {semanticSpatialResultId ? (
        <Button
          size="compact"
          shape="pill"
@@ -901,9 +935,11 @@ export function M4ReplayThreatVisual({
            : playbackController.playback.playing ? "воспроизведение" : "пауза / seek"}
        </small>
      </div>
-      <div>
-        <span>Spatial evidence</span>
-        <strong>{classifiedSpatialLayer
+      {showSpatialOverlaySummary ? (
+        <>
+          <div>
+            <span>Spatial evidence</span>
+            <strong>{classifiedSpatialLayer
          ? classifiedSpatialFrame
            ? replaceClassifiedPointCloud
              ? `${classifiedSpatialFrame.pointsMapGravityLocalXyzM.length.toLocaleString("ru-RU")} TGS points · ${classifiedCellCount.toLocaleString("ru-RU")} cells`
@@ -939,15 +975,15 @@ export function M4ReplayThreatVisual({
                    : showMediaPoints && cameraPointOverlay.error
                      ? " · накопленное camera cloud недоступно"
                      : ""}
-                {semantic && spatialSemanticFrame
+                {semanticSpatialResultId && spatialSemanticFrame
                  ? ` · semantic L ${spatialSemanticFrame.counts.labeled} · A ${spatialSemanticFrame.counts.ambiguous} · U ${spatialSemanticFrame.counts.unprojected} · Ø ${spatialSemanticFrame.counts.absent}`
-                  : semantic ? " · semantic buffer" : ""}
+                  : semanticSpatialResultId ? " · semantic buffer" : ""}
              </>
            )}</small>
-      </div>
-      <div>
-        <span>{classifiedSpatialLayer ? "TGS fail-closed" : "Virtual corridor"}</span>
-        <strong>{classifiedSpatialLayer
+          </div>
+          <div>
+            <span>{classifiedSpatialLayer ? "TGS fail-closed" : "Virtual corridor"}</span>
+            <strong>{classifiedSpatialLayer
          ? classifiedSpatialFrame
            ? `${classifiedCellCounts.occupied} occupied · ${classifiedCellCounts.rejected} rejected · ${classifiedCellCounts.unobserved} unobserved`
            : classifiedSpatialLayer.loading || displayingBufferedFrame ? "loading" : "unavailable"
@@ -957,7 +993,9 @@ export function M4ReplayThreatVisual({
            ? `${classifiedCellCounts.ground} ground-support · visual review only · navigation authority OFF`
            : "visual review only · navigation authority OFF"
          : `${metadata.timeline.corridor.forwardLengthM} м · body ${metadata.timeline.rig.lengthM}×${metadata.timeline.rig.widthM} м · REPLAY-SIMULATED`}</small>
-      </div>
+          </div>
+        </>
+      ) : null}
    </div>
  ) : undefined;

@@ -1153,13 +1191,13 @@ export function M4ReplayThreatVisual({
            <span>{timelineFrame.error}</span>
          </div>
        ) : null}
-        {semantic && semanticTimeline.loading ? (
+        {semanticSpatialResultId && semanticTimeline.loading ? (
          <div className="m4-replay-threat-visual__buffering" role="status">
            <span className="busy-indicator" aria-hidden="true" />
            <span>Догружаем semantic-point evidence E47</span>
          </div>
        ) : null}
-        {semanticTimeline.error ? (
+        {semanticSpatialResultId && semanticTimeline.error ? (
          <div className="m4-replay-threat-visual__buffering" role="alert">
            <Icon name="alert" size={16} />
            <span>{semanticTimeline.error}</span>
@@ -1201,7 +1239,7 @@ export function M4ReplayThreatVisual({
    <div className="l3-visual-audit m4-replay-threat-visual">
      <LaboratoryEvidenceViewer
        label={semantic
-          ? "E47 semantic + SLAM diagnostic replay"
+          ? semantic.label ?? "Semantic diagnostic replay"
          : `${evidenceLabel} recorded-realtime replay`}
        className="m4-replay-threat-evidence-viewer"
        mode={mediaMode ?? "none"}
@@ -4,11 +4,15 @@ import {
  LaboratorySummary,
  LaboratoryWorkTemplate,
 } from "../../components/laboratory/LaboratoryPresentation";
-import type { VegetationShadowResult } from "../../core/laboratory/vegetationShadow";
+import {
+  vegetationVideoMaskUrl,
+  type VegetationShadowResult,
+} from "../../core/laboratory/vegetationShadow";
 import {
  M48MaskComparisonVisual,
  type M48MaskComparisonCase,
 } from "./M48FailureAtlasVisual";
+import { M4ReplayThreatVisual } from "./M4ReplayThreatVisual";

 function decimal(value: number, digits = 1): string {
  return value.toLocaleString("ru-RU", { maximumFractionDigits: digits });
@@ -63,20 +67,28 @@ export function VegetationShadowResultView({
      summary={(
        <LaboratorySummary
          title="LAB V1 · готовые модели растительности"
-          description="Штатный M4.8-инструмент сравнивает две готовые fine-64 модели на полном GOOSE validation split и на 12 truth-backed hard cases, выбранных только по наличию нужной растительности. Sealed evidence открывается локально без Worker 006."
-          status="Truth-backed model comparison · route transfer не принят"
+          description={result.routeVideo
+            ? "M4.8 сохраняет truth-backed сравнение моделей, а штатный M4.7 viewer показывает фактический DDRNet prediction на всей записи RAVNOVES00. Все 4489 масок запечатаны локально и открываются без Worker 006."
+            : "Штатный M4.8-инструмент сравнивает две готовые fine-64 модели на полном GOOSE validation split и на 12 truth-backed hard cases, выбранных только по наличию нужной растительности. Sealed evidence открывается локально без Worker 006."}
+          status={result.routeVideo
+            ? "DDRNet full-video prediction ready · route truth отсутствует"
+            : "Truth-backed model comparison · route transfer не принят"}
          statusTone="warning"
          facts={[
            { label: "Источник", value: "GOOSE validation · 962 размеченных кадра · 12 vegetation hard cases" },
            { label: "Сравнение", value: "DDRNet-39 vs PPLiteSeg · official fine-64 weights" },
            { label: "Кейсы", value: "трава · куст · ствол · крона · изгородь · лес · посевы" },
+            ...(result.routeVideo ? [{
+              label: "Видео",
+              value: "RAVNOVES00 · 4489/4489 DDRNet masks · exact recorded sequence",
+            }] : []),
            { label: "Authority", value: `${rigLabel} · MODEL QUALIFICATION ONLY · commands OFF` },
          ]}
          brief={{
            question: "Какие готовые веса лучше различают проезжаемую траву, кусты и стволы на размеченных off-road кадрах?",
            approach: "Обе модели последовательно прогнаны в одном изолированном CUDA-runtime на 962 кадрах. 12 визуальных кейсов выбраны детерминированно по truth-поддержке восьми растительных классов; один M4.8 viewer показывает source, truth, prediction и material-error для выбранной модели.",
-            principalResult: `${selected.loadedModelName} лидирует по vegetation IoU: ${decimal(selected.vegetationMeanIouPercent, 2)}% против ${decimal(alternative.vegetationMeanIouPercent, 2)}%. Ошибки по каждому типу теперь проверяются в одном штатном инструменте.`,
-            limitation: "Это внешний GOOSE-домен, а не наш fisheye/off-road маршрут. Папоротник отдельным классом отсутствует; RAVNOVES00 не содержит truth-backed vegetation island и не используется как главное визуальное доказательство.",
+            principalResult: `${selected.loadedModelName} лидирует по vegetation IoU: ${decimal(selected.vegetationMeanIouPercent, 2)}% против ${decimal(alternative.vegetationMeanIouPercent, 2)}%. ${result.routeVideo ? "Его фактическая temporal stability теперь видна на всех 4489 кадрах штатного recorded viewer." : "Ошибки по каждому типу проверяются в одном штатном инструменте."}`,
+            limitation: "GOOSE — внешний размеченный домен; RAVNOVES00 — наш fisheye, но без ручной truth-разметки. Full-video слой показывает prediction, а не доказывает правильность. Папоротник отдельным классом отсутствует.",
          }}
          method={{
            completeness: "complete",
@@ -93,17 +105,42 @@ export function VegetationShadowResultView({
        />
      )}
      evidence={(
-        <LaboratoryEvidence
-          eyebrow="M4.8 · GOOSE VEGETATION HARD CASES"
-          title="ERROR: красный — пропуск · жёлтый — лишнее · фиолетовый — перепутан тип · зелёный — совпадение"
-          kind="diagnostic-model"
-          resizable
-        >
-          <M48MaskComparisonVisual
-            cases={comparisonCases(result)}
-            initialCandidate={result.selectedCandidate}
-          />
-        </LaboratoryEvidence>
+        <>
+          <LaboratoryEvidence
+            eyebrow="M4.8 · GOOSE VEGETATION HARD CASES"
+            title="ERROR: красный — пропуск · жёлтый — лишнее · фиолетовый — перепутан тип · зелёный — совпадение"
+            kind="diagnostic-model"
+            resizable
+          >
+            <M48MaskComparisonVisual
+              cases={comparisonCases(result)}
+              initialCandidate={result.selectedCandidate}
+            />
+          </LaboratoryEvidence>
+          {result.routeVideo ? (
+            <LaboratoryEvidence
+              eyebrow="M4.7 · RAVNOVES00 FULL VIDEO"
+              title="DDRNet PREDICTION · 4489/4489 кадров · TRUTH для этой записи отсутствует"
+              kind="diagnostic-model"
+              resizable
+            >
+              <M4ReplayThreatVisual
+                resultId={result.routeVideo.baseM4ResultId}
+                evidenceLabel="LAB V1 · DDRNet"
+                showReferenceMediaLayers={false}
+                showSpatialOverlaySummary={false}
+                semantic={{
+                  resultId: result.routeVideo.workerResultId,
+                  spatialResultId: null,
+                  taxonomy: result.routeVideo.taxonomy,
+                  maskUrl: (sequence) => vegetationVideoMaskUrl(result.resultId, sequence),
+                  label: "DDRNet vegetation prediction · recorded video",
+                  maskAriaLabel: "DDRNet vegetation prediction",
+                }}
+              />
+            </LaboratoryEvidence>
+          ) : null}
+        </>
      )}
      result={(
        <LaboratoryResultSummary
@@ -146,11 +183,16 @@ export function VegetationShadowResultView({
              value: "12 truth-backed cases",
              hint: "8 vegetation strata · Worker для открытия не требуется",
            },
+            ...(result.routeVideo ? [{
+              label: "Route video",
+              value: "4489/4489 masks",
+              hint: "DDRNet prediction · exact sequence · Worker-independent playback",
+            }] : []),
          ]}
          conclusion={{
            proved: "Обе официальные fine-64 модели воспроизводимо запускаются на Worker 006; DDRNet лучше по aggregate vegetation IoU. Truth-backed hard cases прямо показывают траву, кусты и стволы, а не случайные автомобили и здания.",
-            notProved: "Не доказаны accuracy на нашем fisheye-домене, папоротник как отдельный материал, temporal stability, collision safety и physical-live поведение ровера.",
-            decision: "Сохранить DDRNet как стартовый vegetation candidate. Mission-policy и автоматическое переключение пресетов подключать только после truth-backed island нашего офф-роуда; LiDAR/TGS fail-closed геометрию не ослаблять.",
+            notProved: "Не доказаны accuracy на нашем fisheye-домене, папоротник как отдельный материал, collision safety и physical-live поведение ровера. Видео позволяет увидеть temporal stability, но без truth не превращает её в метрику качества.",
+            decision: "Смотреть полный prediction на видео и собирать конкретные temporal/domain failure cases. DDRNet остаётся diagnostic candidate; LiDAR/TGS fail-closed геометрию не ослаблять.",
          }}
        />
      )}