feat(review): implement perception benchmarker and update ignore list

Author: dibsonmuad

.gitignore

@@ -11,6 +11,7 @@ core/target/
 /data/
 *.jsonl
 *.pt
+yolov8*.pt
 
 # Environment
 .env

tools/review/benchmark.py

@@ -0,0 +1,101 @@
+import json
+import numpy as np
+import argparse
+from pathlib import Path
+
+
+def load_jsonl(path):
+    data = []
+    if not Path(path).exists():
+        return data
+    with open(path, 'r') as f:
+        for line in f:
+            data.append(json.loads(line))
+    return data
+
+
+def compute_pck(gt_pose, pred_pose, threshold=0.05):
+    """
+    Computes Percentage of Correct Keypoints (PCK).
+    threshold: normalized distance relative to bbox diagonal.
+    """
+    gt = np.array(gt_pose)
+    pred = np.array(pred_pose)
+    
+    # Simple L2 distance on normalized coordinates
+    distances = np.linalg.norm(gt - pred, axis=1)
+    correct = distances < threshold
+    return np.mean(correct)
+
+
+def run_benchmark(gt_path, pred_path):
+    print(f"[BENCHMARK] Comparing {gt_path} vs {pred_path}...")
+    
+    gt_data = load_jsonl(gt_path)
+    pred_data = load_jsonl(pred_path)
+    
+    if not gt_data or not pred_data:
+        print("[ERROR] Missing input data for benchmarking.")
+        return
+
+    # Filter pred_data to only include player events
+    preds = [p for p in pred_data if p.get("kind") == "player"]
+    
+    stats = {
+        "total_gt_samples": len(gt_data),
+        "matches_found": 0,
+        "id_accuracy": 0.0,
+        "avg_pck": 0.0
+    }
+
+    pck_scores = []
+    id_matches = 0
+
+    for gt in gt_data:
+        t_ms = gt["t_ms"]
+        gt_tid = gt["track_id"]
+        
+        # Find corresponding prediction
+        # Search for same t_ms and closest bbox/id
+        match = None
+        for p in preds:
+            if abs(p["t_ms"] - t_ms) < 33: # Within 1 frame at 30fps
+                # For now, we trust track_id matching for the audit
+                if p["track_id"] == gt_tid:
+                    match = p
+                    break
+        
+        if match:
+            stats["matches_found"] += 1
+            if gt.get("type") == "id_verification":
+                if gt["label"] == "correct":
+                    id_matches += 1
+            
+            # If both have poses, compute PCK
+            # (Requires pose to be in Shush-P JSONL, which we added recently)
+            if "pose_2d" in match and "pose_2d" in gt:
+                pck = compute_pck(gt["pose_2d"], match["pose_2d"])
+                pck_scores.append(pck)
+
+    if stats["matches_found"] > 0:
+        stats["id_accuracy"] = id_matches / stats["matches_found"]
+        if pck_scores:
+            stats["avg_pck"] = np.mean(pck_scores)
+
+    print("\n--- PERCEPTION BENCHMARK REPORT ---")
+    print(f"GT Samples: {stats['total_gt_samples']}")
+    print(f"Matches:    {stats['matches_found']}")
+    print(f"ID Acc:     {stats['id_accuracy']:.2%}")
+    print(f"Avg PCK:    {stats['avg_pck']:.2%}")
+    print("-----------------------------------\n")
+    
+    return stats
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--gt", default="data/training/manual_gt.jsonl")
+    parser.add_argument("--pred", default="data/intelligent_game_dna.jsonl")
+    args = parser.parse_args()
+    
+    run_benchmark(args.gt, args.pred)