feat(ocr-pipeline): ground-truth comparison tool for column detection

Side-by-side view: auto result (readonly) vs GT editor where teacher
draws correct columns. Diff table shows Auto vs GT with IoU matching.
GT data persisted per session for algorithm tuning.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-02-27 22:48:37 +01:00
parent 03fa186fec
commit 587b066a40
5 changed files with 352 additions and 16 deletions

View File

@@ -4,21 +4,44 @@ import { useCallback, useEffect, useState } from 'react'
import type { ColumnResult, ColumnGroundTruth, PageRegion } from '@/app/(admin)/ai/ocr-pipeline/types'
import { ColumnControls } from './ColumnControls'
import { ManualColumnEditor } from './ManualColumnEditor'
import type { ColumnTypeKey } from '@/app/(admin)/ai/ocr-pipeline/types'
const KLAUSUR_API = '/klausur-api'
type ViewMode = 'normal' | 'ground-truth' | 'manual'
interface StepColumnDetectionProps {
sessionId: string | null
onNext: () => void
}
/** Convert PageRegion[] to divider percentages + column types for ManualColumnEditor */
function columnsToEditorState(
columns: PageRegion[],
imageWidth: number
): { dividers: number[]; columnTypes: ColumnTypeKey[] } {
if (!columns.length || !imageWidth) return { dividers: [], columnTypes: [] }
const sorted = [...columns].sort((a, b) => a.x - b.x)
const dividers: number[] = []
const columnTypes: ColumnTypeKey[] = sorted.map(c => c.type)
for (let i = 1; i < sorted.length; i++) {
const xPct = (sorted[i].x / imageWidth) * 100
dividers.push(xPct)
}
return { dividers, columnTypes }
}
export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionProps) {
const [columnResult, setColumnResult] = useState<ColumnResult | null>(null)
const [detecting, setDetecting] = useState(false)
const [error, setError] = useState<string | null>(null)
const [manualMode, setManualMode] = useState(false)
const [viewMode, setViewMode] = useState<ViewMode>('normal')
const [applying, setApplying] = useState(false)
const [imageDimensions, setImageDimensions] = useState<{ width: number; height: number } | null>(null)
const [savedGtColumns, setSavedGtColumns] = useState<PageRegion[] | null>(null)
// Fetch session info (image dimensions) + check for cached column result
useEffect(() => {
@@ -49,6 +72,24 @@ export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionPr
// eslint-disable-next-line react-hooks/exhaustive-deps
}, [sessionId])
// Load saved GT if exists
useEffect(() => {
if (!sessionId) return
const fetchGt = async () => {
try {
const res = await fetch(`${KLAUSUR_API}/api/v1/ocr-pipeline/sessions/${sessionId}/ground-truth/columns`)
if (res.ok) {
const data = await res.json()
const corrected = data.columns_gt?.corrected_columns
if (corrected) setSavedGtColumns(corrected)
}
} catch {
// No saved GT - that's fine
}
}
fetchGt()
}, [sessionId])
const runAutoDetection = useCallback(async () => {
if (!sessionId) return
setDetecting(true)
@@ -106,7 +147,30 @@ export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionPr
columns: data.columns,
duration_seconds: data.duration_seconds ?? 0,
})
setManualMode(false)
setViewMode('normal')
} catch (e) {
setError(e instanceof Error ? e.message : 'Fehler beim Speichern')
} finally {
setApplying(false)
}
}, [sessionId])
const handleGtApply = useCallback(async (columns: PageRegion[]) => {
if (!sessionId) return
setApplying(true)
setError(null)
try {
const gt: ColumnGroundTruth = {
is_correct: false,
corrected_columns: columns,
}
await fetch(`${KLAUSUR_API}/api/v1/ocr-pipeline/sessions/${sessionId}/ground-truth/columns`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify(gt),
})
setSavedGtColumns(columns)
setViewMode('normal')
} catch (e) {
setError(e instanceof Error ? e.message : 'Fehler beim Speichern')
} finally {
@@ -131,6 +195,11 @@ export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionPr
const dewarpedUrl = `${KLAUSUR_API}/api/v1/ocr-pipeline/sessions/${sessionId}/image/dewarped`
const overlayUrl = `${KLAUSUR_API}/api/v1/ocr-pipeline/sessions/${sessionId}/image/columns-overlay`
// Pre-compute editor state from saved GT or auto columns for GT mode
const gtInitial = savedGtColumns
? columnsToEditorState(savedGtColumns, imageDimensions?.width ?? 1000)
: undefined
return (
<div className="space-y-4">
{/* Loading indicator */}
@@ -141,18 +210,77 @@ export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionPr
</div>
)}
{manualMode ? (
/* Manual column editor */
{viewMode === 'manual' ? (
/* Manual column editor - overwrites column_result */
<ManualColumnEditor
imageUrl={dewarpedUrl}
imageWidth={imageDimensions?.width ?? 1000}
imageHeight={imageDimensions?.height ?? 1400}
onApply={handleManualApply}
onCancel={() => setManualMode(false)}
onCancel={() => setViewMode('normal')}
applying={applying}
mode="manual"
/>
) : viewMode === 'ground-truth' ? (
/* GT mode: auto result (left, readonly) + GT editor (right) */
<div className="grid grid-cols-2 gap-4">
{/* Left: Auto result (readonly overlay) */}
<div>
<div className="text-xs font-medium text-gray-500 dark:text-gray-400 mb-1">
Auto-Ergebnis (readonly)
</div>
<div className="border rounded-lg overflow-hidden dark:border-gray-700 bg-gray-50 dark:bg-gray-900">
{columnResult ? (
// eslint-disable-next-line @next/next/no-img-element
<img
src={`${overlayUrl}?t=${Date.now()}`}
alt="Auto Spalten-Overlay"
className="w-full h-auto"
/>
) : (
<div className="aspect-[3/4] flex items-center justify-center text-gray-400 text-sm">
Keine Auto-Daten
</div>
)}
</div>
{/* Auto column list */}
{columnResult && (
<div className="mt-2 space-y-1">
<div className="text-xs font-medium text-gray-500 dark:text-gray-400">
Auto: {columnResult.columns.length} Spalten
</div>
{columnResult.columns
.filter(c => c.type.startsWith('column') || c.type === 'page_ref')
.map((col, i) => (
<div key={i} className="text-xs text-gray-500 dark:text-gray-400 font-mono">
{i + 1}. {col.type} x={col.x} w={col.width}
</div>
))}
</div>
)}
</div>
{/* Right: GT editor */}
<div>
<div className="text-xs font-medium text-gray-500 dark:text-gray-400 mb-1">
Ground Truth Editor
</div>
<ManualColumnEditor
imageUrl={dewarpedUrl}
imageWidth={imageDimensions?.width ?? 1000}
imageHeight={imageDimensions?.height ?? 1400}
onApply={handleGtApply}
onCancel={() => setViewMode('normal')}
applying={applying}
mode="ground-truth"
layout="stacked"
initialDividers={gtInitial?.dividers}
initialColumnTypes={gtInitial?.columnTypes}
/>
</div>
</div>
) : (
/* Image comparison: overlay (left) vs clean (right) */
/* Normal mode: overlay (left) vs clean (right) */
<div className="grid grid-cols-2 gap-4">
<div>
<div className="text-xs font-medium text-gray-500 dark:text-gray-400 mb-1">
@@ -190,14 +318,16 @@ export function StepColumnDetection({ sessionId, onNext }: StepColumnDetectionPr
)}
{/* Controls */}
{!manualMode && (
{viewMode === 'normal' && (
<ColumnControls
columnResult={columnResult}
onRerun={handleRerun}
onManualMode={() => setManualMode(true)}
onManualMode={() => setViewMode('manual')}
onGtMode={() => setViewMode('ground-truth')}
onGroundTruth={handleGroundTruth}
onNext={onNext}
isDetecting={detecting}
savedGtColumns={savedGtColumns}
/>
)}