Upload segmentation/pipeline_60s_run2/scripts/phase6_fuse.py with huggingface_hub

segmentation/pipeline_60s_run2/scripts/phase6_fuse.py

@@ -0,0 +1,94 @@
+import csv
+import glob
+import json
+
+import cv2
+import numpy as np
+
+ROOT = "/root/bdmc_pipeline"
+H, W = 1920, 1080
+fx = fy = H / (2 * np.tan(np.radians(72.0 / 2)))
+K = np.array([[fx, 0, W / 2], [0, fy, H / 2], [0, 0, 1.0]])
+RADIUS_PX = 110
+
+devmaps = np.load(f"{ROOT}/outputs/devmaps_small.npz")["dev"].astype(np.float32)
+devmaps_m = (devmaps - 128.0) / 1000.0
+defects = list(csv.DictReader(open(f"{ROOT}/outputs/defects.csv")))
+wins = {}
+for p in sorted(glob.glob(f"{ROOT}/outputs/tracks/win_*.npz")):
+    z = np.load(p)
+    wins[int(z["start"])] = {"tracks": z["tracks"], "vis": z["vis"].reshape(z["vis"].shape[0], -1)}
+
+tri_meta = {}
+for p in sorted(glob.glob(f"{ROOT}/outputs/tracks/triangulated_*.npz")):
+    z = np.load(p)
+    tri_meta[int(z["start"])] = {"xyz": z["xyz"], "Rs": z["Rs"], "ts": z["ts"],
+                                 "rmse_med": float(np.median(z["rmse"]))}
+
+def track_support(fst, dur, cxc, cyc):
+    dvs = []
+    for s, w in wins.items():
+        lo, hi = max(fst, s), min(fst + dur - 1, s + w["tracks"].shape[0] - 1)
+        for f in range(lo, hi + 1):
+            tr = w["tracks"][f - s]
+            v = w["vis"][f - s].astype(bool)
+            xi = np.clip(tr[:, 0].round().astype(int), 0, W - 1)
+            yi = np.clip(tr[:, 1].round().astype(int), 0, H - 1)
+            dist = np.hypot(xi - cxc, yi - cyc)
+            near = v & (dist < RADIUS_PX)
+            if near.sum() == 0:
+                continue
+            dvs.append(devmaps_m[f - 1][yi[near] // 4, xi[near] // 4])
+    if not dvs:
+        return None
+    dv = np.concatenate(dvs)
+    return {"n_obs": int(len(dv)),
+            "support_frac_lt3cm": round(float((dv < -0.03).mean()), 3),
+            "median_dev_cm": round(float(np.median(dv)) * 100, 2),
+            "p25_dev_cm": round(float(np.percentile(dv, 25)) * 100, 2)}
+
+rows_out = []
+for r in defects:
+    c = str(r["centroid_px"]).replace("[", "").replace("]", "").replace(",", " ").split()
+    cxc, cyc = float(c[0]), float(c[1])
+    fst, dur = int(r["first_frame"]), int(r["duration_frames"])
+    rec = dict(r)
+    ts = track_support(fst, dur, cxc, cyc)
+    if ts is None or ts["n_obs"] < 5:
+        rec.update({"track_n_obs": ts["n_obs"] if ts else 0, "track_median_dev_cm": None,
+                    "track_support_frac": None, "confidence": "no_coverage"})
+    else:
+        med, sf = ts["median_dev_cm"], ts["support_frac_lt3cm"]
+        conf = ("agree" if (med <= -3.0 and sf >= 0.4)
+                else "partial" if med <= -1.5 else "disagree")
+        rec.update({**ts, "track_support_frac": sf, "track_median_dev_cm": med,
+                    "confidence": conf})
+    rows_out.append(rec)
+
+cols = ["first_frame", "t_sec", "duration_frames", "duration_s", "peak_dev_cm", "area_px",
+        "centroid_px", "n_obs", "median_dev_cm", "p25_dev_cm", "support_frac_lt3cm",
+        "confidence"]
+with open(f"{ROOT}/outputs/defects_fused.csv", "w", newline="") as fh:
+    wr = csv.DictWriter(fh, fieldnames=cols)
+    wr.writeheader()
+    wr.writerows([{k: r.get(k) for k in cols} for r in rows_out])
+
+tri_files = glob.glob(f"{ROOT}/outputs/tracks/triangulated_*.npz")
+if tri_files:
+    med_rmse = float(np.median(np.concatenate([np.load(p)["rmse"] for p in tri_files])))
+else:
+    med_rmse = 0.0
+
+report = {
+    "n_defects": len(rows_out),
+    "confidence_counts": {c: sum(1 for r in rows_out if r.get("confidence") == c)
+                          for c in ["agree", "partial", "disagree", "no_coverage"]},
+    "method": "per-defect: gather all CoTracker3 observations passing through defect region during its confirmed lifetime; compare their depth-residual samples against the flag",
+    "median_triangulated_rmse_px": round(med_rmse, 2),
+    "sigma_proxy_triangulated_cm_at_5m": round(float(med_rmse * 5.0 / fx * 100), 2),
+    "depth_tile_sigma_cm": json.load(open(f"{ROOT}/outputs/phase3_summary.json"))["tile_sigma_median_cm"],
+    "note": "no GPS/IMU: egomotion estimated visually (Kabsch on depth-anchored correspondences)",
+}
+json.dump(report, open(f"{ROOT}/outputs/phase6_report.json", "w"), indent=1)
+print(json.dumps(report, indent=1))
+print("Phase 6 fusion complete", flush=True)