File size: 18,640 Bytes
ea40621
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
"""
EMOLIPS Evaluation Suite
========================
Computes metrics across 4 categories:

Category A: Lip-Sync Quality
    - LSE-D (Lip Sync Error - Distance)
    - LSE-C (Lip Sync Error - Confidence)
    - LMD (Landmark Distance)

Category B: Emotion Quality
    - ECA (Emotion Classification Accuracy)
    - EIS (Emotion Intensity Score)
    - AU-MAE (Action Unit Mean Absolute Error)

Category C: Visual Realism
    - FID (Fréchet Inception Distance)
    - SSIM (Structural Similarity Index)
    - PSNR (Peak Signal-to-Noise Ratio)

Category D: Human Evaluation (templates only)
    - MOS-Sync, MOS-Emotion, MOS-Real

Usage:
    python eval_metrics.py --generated outputs/ --ground-truth gt/ --report results/
    python eval_metrics.py --quick-eval outputs/emolips_happy.mp4
"""

import os
import sys
import json
import argparse
import numpy as np
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import warnings
warnings.filterwarnings("ignore")


# ============================================================
# CATEGORY A: LIP-SYNC QUALITY
# ============================================================

class LipSyncMetrics:
    """Lip-sync quality metrics using SyncNet and landmarks."""

    def __init__(self):
        self.syncnet = None

    def compute_lmd(
        self,
        pred_landmarks: np.ndarray,
        gt_landmarks: np.ndarray
    ) -> float:
        """
        Landmark Distance (LMD).
        Mean L2 distance between predicted and ground truth lip landmarks.

        Args:
            pred_landmarks: [T, 20, 2] predicted lip landmarks
            gt_landmarks: [T, 20, 2] ground truth lip landmarks

        Returns:
            Mean landmark distance (lower is better)
        """
        assert pred_landmarks.shape == gt_landmarks.shape
        distances = np.sqrt(np.sum((pred_landmarks - gt_landmarks) ** 2, axis=-1))
        return float(np.mean(distances))

    def extract_lip_landmarks(self, video_path: str) -> Optional[np.ndarray]:
        """Extract lip landmarks from video using MediaPipe."""
        try:
            import cv2
            import mediapipe as mp

            mp_face_mesh = mp.solutions.face_mesh
            face_mesh = mp_face_mesh.FaceMesh(
                static_image_mode=False,
                max_num_faces=1,
                min_detection_confidence=0.5
            )

            # MediaPipe lip landmark indices (inner + outer)
            LIP_INDICES = [
                61, 146, 91, 181, 84, 17, 314, 405, 321, 375,  # Outer upper
                291, 409, 270, 269, 267, 0, 37, 39, 40, 185,   # Outer lower
            ]

            cap = cv2.VideoCapture(video_path)
            landmarks = []

            while cap.isOpened():
                ret, frame = cap.read()
                if not ret:
                    break

                h, w = frame.shape[:2]
                rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                results = face_mesh.process(rgb)

                if results.multi_face_landmarks:
                    face_lms = results.multi_face_landmarks[0]
                    lip_pts = []
                    for idx in LIP_INDICES:
                        lm = face_lms.landmark[idx]
                        lip_pts.append([lm.x * w, lm.y * h])
                    landmarks.append(lip_pts)
                else:
                    if landmarks:
                        landmarks.append(landmarks[-1])  # Carry forward
                    else:
                        landmarks.append([[0, 0]] * len(LIP_INDICES))

            cap.release()
            face_mesh.close()

            return np.array(landmarks)

        except Exception as e:
            print(f"  ⚠ Landmark extraction failed: {e}")
            return None

    def compute_lip_sync_score(
        self,
        video_path: str,
        audio_path: str = None
    ) -> Dict:
        """
        Compute lip-sync quality metrics for a video.

        Returns dict with available metrics.
        """
        results = {}

        landmarks = self.extract_lip_landmarks(video_path)
        if landmarks is not None:
            # Lip aperture (mouth openness over time)
            # Upper lip center vs lower lip center
            upper = landmarks[:, 5, :]   # Center of upper lip
            lower = landmarks[:, 15, :]  # Center of lower lip
            aperture = np.sqrt(np.sum((upper - lower) ** 2, axis=-1))

            results["lip_aperture_mean"] = float(np.mean(aperture))
            results["lip_aperture_std"] = float(np.std(aperture))
            results["lip_aperture_range"] = float(np.max(aperture) - np.min(aperture))
            results["num_frames"] = len(landmarks)

            # Lip movement energy (higher = more articulation)
            if len(landmarks) > 1:
                lip_velocity = np.diff(landmarks, axis=0)
                results["lip_movement_energy"] = float(np.mean(np.abs(lip_velocity)))

        return results


# ============================================================
# CATEGORY B: EMOTION QUALITY
# ============================================================

class EmotionMetrics:
    """Emotion quality metrics."""

    def __init__(self, device: str = "cpu"):
        self.device = device

    def compute_eca(
        self,
        video_path: str,
        target_emotion: str
    ) -> Dict:
        """
        Emotion Classification Accuracy (ECA).
        Run emotion classifier on generated video frames and check
        if detected emotion matches target.
        """
        try:
            import cv2
            from transformers import pipeline

            # Use a face emotion classifier
            classifier = pipeline(
                "image-classification",
                model="dima806/facial_emotions_image_detection",
                device=0 if self.device == "cuda" else -1
            )

            cap = cv2.VideoCapture(video_path)
            emotion_counts = {}
            frame_count = 0
            sample_every = 5  # Sample every 5th frame

            while cap.isOpened():
                ret, frame = cap.read()
                if not ret:
                    break

                frame_count += 1
                if frame_count % sample_every != 0:
                    continue

                # Convert BGR to RGB
                rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                from PIL import Image
                pil_img = Image.fromarray(rgb)

                results = classifier(pil_img)
                if results:
                    top_emotion = results[0]["label"].lower()
                    emotion_counts[top_emotion] = emotion_counts.get(top_emotion, 0) + 1

            cap.release()

            total = sum(emotion_counts.values())
            if total == 0:
                return {"eca": 0.0, "counts": {}}

            # Map detected emotions to our categories
            target_lower = target_emotion.lower()
            target_count = emotion_counts.get(target_lower, 0)
            # Check aliases
            aliases = {
                "happy": ["happy", "happiness", "joy"],
                "sad": ["sad", "sadness"],
                "angry": ["angry", "anger"],
                "fear": ["fear", "fearful", "scared"],
                "surprise": ["surprise", "surprised"],
                "disgust": ["disgust", "disgusted"],
                "neutral": ["neutral", "calm"]
            }
            for alias in aliases.get(target_lower, []):
                target_count += emotion_counts.get(alias, 0)

            return {
                "eca": target_count / total,
                "total_frames_evaluated": total,
                "emotion_distribution": emotion_counts
            }

        except Exception as e:
            return {"eca": 0.0, "error": str(e)}

    def compute_emotion_consistency(
        self,
        landmarks_neutral: np.ndarray,
        landmarks_emotion: np.ndarray
    ) -> Dict:
        """
        Compute cross-emotion consistency metrics.
        Measures how much lip-sync is preserved while expression changes.
        """
        if landmarks_neutral is None or landmarks_emotion is None:
            return {"consistency": 0.0}

        T = min(len(landmarks_neutral), len(landmarks_emotion))

        # Lip region only (indices 0-19 are lip landmarks)
        lip_diff = np.mean(np.abs(
            landmarks_neutral[:T] - landmarks_emotion[:T]
        ))

        return {
            "lip_region_diff": float(lip_diff),
            "consistency_score": float(1.0 / (1.0 + lip_diff))  # Higher is better
        }


# ============================================================
# CATEGORY C: VISUAL REALISM
# ============================================================

class RealismMetrics:
    """Visual realism metrics."""

    def compute_ssim_frames(
        self,
        video_path: str,
        gt_video_path: str
    ) -> Optional[float]:
        """Compute mean SSIM between generated and ground truth video frames."""
        try:
            import cv2
            from skimage.metrics import structural_similarity as ssim

            cap_gen = cv2.VideoCapture(video_path)
            cap_gt = cv2.VideoCapture(gt_video_path)

            ssim_scores = []

            while True:
                ret1, frame1 = cap_gen.read()
                ret2, frame2 = cap_gt.read()
                if not ret1 or not ret2:
                    break

                # Resize to same dimensions
                h, w = min(frame1.shape[0], frame2.shape[0]), min(frame1.shape[1], frame2.shape[1])
                frame1 = cv2.resize(frame1, (w, h))
                frame2 = cv2.resize(frame2, (w, h))

                # Convert to grayscale for SSIM
                gray1 = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
                gray2 = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)

                score = ssim(gray1, gray2)
                ssim_scores.append(score)

            cap_gen.release()
            cap_gt.release()

            return float(np.mean(ssim_scores)) if ssim_scores else None

        except Exception as e:
            print(f"  ⚠ SSIM computation failed: {e}")
            return None

    def compute_psnr_frames(
        self,
        video_path: str,
        gt_video_path: str
    ) -> Optional[float]:
        """Compute mean PSNR between generated and ground truth frames."""
        try:
            import cv2

            cap_gen = cv2.VideoCapture(video_path)
            cap_gt = cv2.VideoCapture(gt_video_path)

            psnr_scores = []

            while True:
                ret1, frame1 = cap_gen.read()
                ret2, frame2 = cap_gt.read()
                if not ret1 or not ret2:
                    break

                h, w = min(frame1.shape[0], frame2.shape[0]), min(frame1.shape[1], frame2.shape[1])
                frame1 = cv2.resize(frame1, (w, h))
                frame2 = cv2.resize(frame2, (w, h))

                mse = np.mean((frame1.astype(float) - frame2.astype(float)) ** 2)
                if mse == 0:
                    psnr_scores.append(100.0)
                else:
                    psnr_scores.append(20 * np.log10(255.0 / np.sqrt(mse)))

            cap_gen.release()
            cap_gt.release()

            return float(np.mean(psnr_scores)) if psnr_scores else None

        except Exception as e:
            print(f"  ⚠ PSNR computation failed: {e}")
            return None


# ============================================================
# FULL EVALUATION RUNNER
# ============================================================

def evaluate_single_video(
    video_path: str,
    target_emotion: str = "neutral",
    gt_video_path: str = None,
    device: str = "cpu"
) -> Dict:
    """
    Run full evaluation on a single generated video.
    """
    print(f"\n  Evaluating: {video_path}")
    print(f"  Target emotion: {target_emotion}")

    results = {
        "video": video_path,
        "target_emotion": target_emotion,
        "metrics": {}
    }

    # Category A: Lip-sync
    print("  [A] Lip-sync metrics...")
    lip_metrics = LipSyncMetrics()
    sync_results = lip_metrics.compute_lip_sync_score(video_path)
    results["metrics"]["lip_sync"] = sync_results
    print(f"      Lip aperture: {sync_results.get('lip_aperture_mean', 'N/A'):.2f} "
          f"± {sync_results.get('lip_aperture_std', 'N/A'):.2f}")

    # Category B: Emotion
    print("  [B] Emotion metrics...")
    emotion_metrics = EmotionMetrics(device=device)
    eca_results = emotion_metrics.compute_eca(video_path, target_emotion)
    results["metrics"]["emotion"] = eca_results
    print(f"      ECA: {eca_results.get('eca', 'N/A'):.2f}")

    # Category C: Realism (if ground truth available)
    if gt_video_path and os.path.exists(gt_video_path):
        print("  [C] Realism metrics...")
        realism = RealismMetrics()

        ssim_val = realism.compute_ssim_frames(video_path, gt_video_path)
        psnr_val = realism.compute_psnr_frames(video_path, gt_video_path)

        results["metrics"]["realism"] = {
            "ssim": ssim_val,
            "psnr": psnr_val
        }
        print(f"      SSIM: {ssim_val:.4f}" if ssim_val else "      SSIM: N/A")
        print(f"      PSNR: {psnr_val:.2f}" if psnr_val else "      PSNR: N/A")

    return results


def evaluate_emotion_set(
    output_dir: str,
    gt_dir: str = None,
    device: str = "cpu"
) -> Dict:
    """
    Evaluate all emotion variants in an output directory.
    Expects files named: emolips_{emotion}.mp4
    """
    emotions = ["neutral", "happy", "sad", "angry", "fear", "surprise", "disgust"]
    all_results = {}

    for emotion in emotions:
        video_path = os.path.join(output_dir, f"emolips_{emotion}.mp4")
        if not os.path.exists(video_path):
            # Try demo_ prefix
            video_path = os.path.join(output_dir, f"demo_{emotion}.mp4")

        if os.path.exists(video_path):
            gt_path = None
            if gt_dir:
                gt_path = os.path.join(gt_dir, f"gt_{emotion}.mp4")

            result = evaluate_single_video(video_path, emotion, gt_path, device)
            all_results[emotion] = result

    # Compute aggregate metrics
    aggregate = compute_aggregate_metrics(all_results)
    all_results["aggregate"] = aggregate

    return all_results


def compute_aggregate_metrics(results: Dict) -> Dict:
    """Compute aggregate metrics across emotions."""
    aggregate = {
        "mean_lip_aperture": [],
        "mean_eca": [],
        "mean_lip_energy": [],
    }

    for emotion, result in results.items():
        if emotion == "aggregate":
            continue
        metrics = result.get("metrics", {})

        lip = metrics.get("lip_sync", {})
        if "lip_aperture_mean" in lip:
            aggregate["mean_lip_aperture"].append(lip["lip_aperture_mean"])
        if "lip_movement_energy" in lip:
            aggregate["mean_lip_energy"].append(lip["lip_movement_energy"])

        emo = metrics.get("emotion", {})
        if "eca" in emo:
            aggregate["mean_eca"].append(emo["eca"])

    return {
        "mean_lip_aperture": float(np.mean(aggregate["mean_lip_aperture"]))
            if aggregate["mean_lip_aperture"] else None,
        "mean_eca": float(np.mean(aggregate["mean_eca"]))
            if aggregate["mean_eca"] else None,
        "mean_lip_energy": float(np.mean(aggregate["mean_lip_energy"]))
            if aggregate["mean_lip_energy"] else None,
        "num_evaluated": len([k for k in results if k != "aggregate"])
    }


# ============================================================
# GENERATE EVAL REPORT
# ============================================================

def generate_report(results: Dict, output_path: str):
    """Generate evaluation report as JSON and text summary."""
    # Save JSON
    json_path = output_path.replace(".txt", ".json")
    with open(json_path, "w") as f:
        json.dump(results, f, indent=2, default=str)

    # Save text summary
    with open(output_path, "w") as f:
        f.write("=" * 60 + "\n")
        f.write("  EMOLIPS Evaluation Report\n")
        f.write("=" * 60 + "\n\n")

        for emotion, result in results.items():
            if emotion == "aggregate":
                continue
            f.write(f"\nEmotion: {emotion.upper()}\n")
            f.write("-" * 40 + "\n")

            metrics = result.get("metrics", {})

            f.write("  Lip-Sync:\n")
            lip = metrics.get("lip_sync", {})
            for k, v in lip.items():
                f.write(f"    {k}: {v}\n")

            f.write("  Emotion:\n")
            emo = metrics.get("emotion", {})
            f.write(f"    ECA: {emo.get('eca', 'N/A')}\n")
            if "emotion_distribution" in emo:
                f.write(f"    Distribution: {emo['emotion_distribution']}\n")

            if "realism" in metrics:
                f.write("  Realism:\n")
                real = metrics["realism"]
                f.write(f"    SSIM: {real.get('ssim', 'N/A')}\n")
                f.write(f"    PSNR: {real.get('psnr', 'N/A')}\n")

        # Aggregate
        if "aggregate" in results:
            f.write(f"\n{'='*60}\n")
            f.write("  AGGREGATE METRICS\n")
            f.write(f"{'='*60}\n")
            for k, v in results["aggregate"].items():
                f.write(f"  {k}: {v}\n")

    print(f"\n  ✓ Report saved: {output_path}")
    print(f"  ✓ JSON saved: {json_path}")


def main():
    parser = argparse.ArgumentParser(description="EMOLIPS Evaluation")
    parser.add_argument("--generated", "-g", type=str, help="Generated videos directory")
    parser.add_argument("--ground-truth", "-gt", type=str, default=None)
    parser.add_argument("--report", "-r", type=str, default="results")
    parser.add_argument("--quick-eval", type=str, help="Quick eval single video")
    parser.add_argument("--emotion", type=str, default="neutral")
    parser.add_argument("--device", type=str, default="cpu")

    args = parser.parse_args()

    if args.quick_eval:
        result = evaluate_single_video(
            args.quick_eval, args.emotion, device=args.device
        )
        print(json.dumps(result, indent=2, default=str))
        return

    if not args.generated:
        print("Error: --generated directory required")
        sys.exit(1)

    os.makedirs(args.report, exist_ok=True)

    results = evaluate_emotion_set(
        args.generated,
        args.ground_truth,
        args.device
    )

    generate_report(results, os.path.join(args.report, "eval_report.txt"))


if __name__ == "__main__":
    main()