Blame for tracker.py

"""
SAM 2 Robust Object Tracker
============================
Architecture:
  - Grounding DINO  → detects WHAT to track (frame 0 + recovery)
  - SAM 2 Video     → tracks WHERE the object is across frames
  - Sliding Window  → processes video in chunks so SAM 2 memory stays accurate
  - Recovery Loop   → if an object disappears, DINO relocalizes it

Output: Bounding Box + Label + ID per frame (no mask overlay, fast + low VRAM)
"""

import os
import cv2
import numpy as np
import torch
from PIL import Image
from typing import Optional


# ──────────────────────────────────────────────────────────
# Constants
# ──────────────────────────────────────────────────────────
MIN_MASK_AREA      = 64          # pixels² — mask smaller than this = "lost"
CHUNK_SIZE_DEFAULT = 120         # frames per SAM-2 sliding window
HOME_DIR           = os.path.expanduser("~")
SAM2_CKPT_DEFAULT  = os.path.join(
    HOME_DIR, ".cache", "torch", "hub", "checkpoints", "sam2.1_hiera_small.pt"
)
SAM2_CFG_DEFAULT   = "configs/sam2.1/sam2.1_hiera_s.yaml"


# ──────────────────────────────────────────────────────────
# TrackedObject — stores per-object state between chunks
# ──────────────────────────────────────────────────────────
class TrackedObject:
    """Holds identity and last known bounding box for one tracked object."""

    def __init__(self, obj_id: int, label: str, box: np.ndarray):
        self.obj_id = obj_id
        self.label  = label
        self.box    = box.astype(np.float32)   # [x1, y1, x2, y2]
        self.lost   = False                    # True if disappeared last chunk
        self.lost_frames = 0                   # consecutive frames without mask

    def __repr__(self):
        return f"TrackedObject(id={self.obj_id}, label='{self.label}', lost={self.lost})"


# ──────────────────────────────────────────────────────────
# VideoFrameStore — thin wrapper around a frames directory
# ──────────────────────────────────────────────────────────
class VideoFrameStore:
    """Extracts video frames to disk with optional stabilization, blur filter, resize."""

    def __init__(self, video_path: str, output_dir: str,
                 target_fps: Optional[float] = None,
                 max_size: int = 720,
                 blur_threshold: float = 0.0,
                 stabilize: bool = False):

        self.video_path      = video_path
        self.output_dir      = output_dir
        self.target_fps      = target_fps
        self.max_size        = max_size
        self.blur_threshold  = blur_threshold
        self.stabilize       = stabilize

        self.frame_paths: list[str] = []  # sorted list of extracted frame paths
        self.orig_fps  = 0.0
        self.width     = 0
        self.height    = 0

    # ------------------------------------------------------------------
    def extract(self) -> int:
        """Run extraction. Returns number of frames saved."""
        import shutil
        if os.path.exists(self.output_dir):
            shutil.rmtree(self.output_dir)
        os.makedirs(self.output_dir)

        cap = cv2.VideoCapture(self.video_path)
        if not cap.isOpened():
            raise RuntimeError(f"Cannot open video: {self.video_path}")

        self.orig_fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
        raw_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        raw_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

        # Compute output resolution (keep aspect ratio)
        scale = min(1.0, self.max_size / max(raw_w, raw_h))
        self.width  = int(raw_w * scale)
        self.height = int(raw_h * scale)

        # How many original frames to skip between each saved frame
        sample_interval = max(1, round(self.orig_fps / self.target_fps)) \
            if self.target_fps and self.target_fps > 0 else 1

        stab_diff = None
        if self.stabilize:
            stab_diff = self._compute_stabilization(cap, raw_w, raw_h)
            cap.release()
            cap = cv2.VideoCapture(self.video_path)

        saved = 0
        orig_idx = 0

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

            # --- sample at target fps ---
            if orig_idx % sample_interval != 0:
                orig_idx += 1
                continue

            # --- apply stabilization warp ---
            if stab_diff is not None and orig_idx < len(stab_diff):
                dx, dy, da = stab_diff[orig_idx]
                M = np.array([[np.cos(da), -np.sin(da), dx],
                               [np.sin(da),  np.cos(da), dy]], dtype=np.float32)
                frame = cv2.warpAffine(frame, M, (raw_w, raw_h))

            # --- resize ---
            if scale < 1.0:
                frame = cv2.resize(frame, (self.width, self.height), interpolation=cv2.INTER_AREA)

            # --- blur filter ---
            if self.blur_threshold > 0:
                gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
                if cv2.Laplacian(gray, cv2.CV_64F).var() < self.blur_threshold:
                    orig_idx += 1
                    continue

            path = os.path.join(self.output_dir, f"{saved:05d}.jpg")
            cv2.imwrite(path, frame, [cv2.IMWRITE_JPEG_QUALITY, 95])
            self.frame_paths.append(path)
            saved += 1
            orig_idx += 1

        cap.release()

        # Fallback: if blur filter ate everything, save at least 1 frame
        if saved == 0:
            print("[WARN] All frames were blurry — saving 1 raw frame as fallback.")
            cap = cv2.VideoCapture(self.video_path)
            ret, frame = cap.read()
            cap.release()
            if ret:
                if scale < 1.0:
                    frame = cv2.resize(frame, (self.width, self.height))
                path = os.path.join(self.output_dir, "00000.jpg")
                cv2.imwrite(path, frame)
                self.frame_paths.append(path)
                saved = 1

        print(f"[Extract] Saved {saved} frames → {self.output_dir}")
        return saved

    # ------------------------------------------------------------------
    def _compute_stabilization(self, cap, raw_w, raw_h):
        """ORB-based motion estimation → smoothed correction matrix per frame."""
        print("[Stabilize] Computing ORB motion trajectory …")
        transforms = []
        prev_gray  = None
        scale      = 480.0 / max(raw_w, raw_h)

        while True:
            ret, frame = cap.read()
            if not ret:
                break
            small = cv2.resize(frame, (int(raw_w * scale), int(raw_h * scale)))
            gray  = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY)

            dx = dy = da = 0.0
            if prev_gray is not None:
                orb = cv2.ORB_create(300)
                kp1, d1 = orb.detectAndCompute(prev_gray, None)
                kp2, d2 = orb.detectAndCompute(gray, None)
                if d1 is not None and d2 is not None and len(kp1) > 5 and len(kp2) > 5:
                    bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
                    matches = sorted(bf.match(d1, d2), key=lambda m: m.distance)[:50]
                    if len(matches) >= 4:
                        pts1 = np.float32([kp1[m.queryIdx].pt for m in matches])
                        pts2 = np.float32([kp2[m.trainIdx].pt for m in matches])
                        M, _ = cv2.estimateAffinePartial2D(pts1, pts2)
                        if M is not None:
                            dx = M[0, 2] / scale
                            dy = M[1, 2] / scale
                            da = np.arctan2(M[1, 0], M[0, 0])
            transforms.append(np.array([dx, dy, da]))
            prev_gray = gray

        transforms = np.array(transforms)
        traj       = np.cumsum(transforms, axis=0)
        radius     = max(1, min(30, len(traj) // 2))
        smooth     = np.copy(traj)
        for i in range(len(traj)):
            s, e    = max(0, i - radius), min(len(traj), i + radius + 1)
            smooth[i] = np.mean(traj[s:e], axis=0)
        return smooth - traj  # correction per frame


# ──────────────────────────────────────────────────────────
# DinoDetector — wraps Grounding DINO for prompt detection
# ──────────────────────────────────────────────────────────
class DinoDetector:
    """Loads Grounding DINO and runs chunked prompt detection with NMS."""

    CHUNK_SIZE = 15   # max vocabulary items per DINO call (avoids token overflow)

    def __init__(self, device: torch.device):
        self.device = device
        self.processor = None
        self.model     = None

    def load(self):
        from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
        print("[DINO] Loading Grounding DINO Base …")
        self.processor = AutoProcessor.from_pretrained("IDEA-Research/grounding-dino-base")
        self.model     = AutoModelForZeroShotObjectDetection.from_pretrained(
            "IDEA-Research/grounding-dino-base"
        ).to(self.device).eval()
        print("[DINO] Loaded.")

    # ------------------------------------------------------------------
    def detect(self, image_path: str, prompt: str,
               box_threshold: float = 0.30,
               text_threshold: float = 0.25,
               iou_threshold:  float = 0.45
               ) -> tuple[np.ndarray, np.ndarray, list[str]]:
        """
        Returns (boxes [N,4], scores [N], labels [N]) in pixel coords.
        Prompt can be a multi-line string with # comments and comma-separated items.
        """
        image_pil = Image.open(image_path).convert("RGB")
        items     = self._parse_prompt(prompt)

        if not items:
            return np.empty((0, 4)), np.array([]), []

        # Split vocabulary into chunks of CHUNK_SIZE
        chunks = [items[i:i+self.CHUNK_SIZE]
                  for i in range(0, len(items), self.CHUNK_SIZE)]

        all_boxes, all_scores, all_labels = [], [], []

        for idx, chunk in enumerate(chunks):
            chunk_text = " . ".join(chunk) + " ."
            print(f"  [DINO] chunk {idx+1}/{len(chunks)}: {chunk_text[:80]}…")
            inputs = self.processor(
                images=image_pil, text=chunk_text, return_tensors="pt"
            ).to(self.device)

            with torch.no_grad():
                outputs = self.model(**inputs)

            results = self._post_process(outputs, inputs.input_ids, image_pil,
                                         box_threshold, text_threshold)
            boxes  = results["boxes"].cpu().numpy()
            scores = results["scores"].cpu().numpy()
            labels = results["labels"]

            all_boxes.extend(boxes)
            all_scores.extend(scores)
            all_labels.extend(labels)

        if not all_boxes:
            return np.empty((0, 4)), np.array([]), []

        all_boxes  = np.array(all_boxes)
        all_scores = np.array(all_scores)

        keep = self._nms(all_boxes, all_scores, iou_threshold)
        return all_boxes[keep], all_scores[keep], [all_labels[k] for k in keep]

    # ------------------------------------------------------------------
    def _post_process(self, outputs, input_ids, image_pil, box_thresh, text_thresh):
        try:
            return self.processor.post_process_grounded_object_detection(
                outputs, input_ids,
                box_threshold=box_thresh, text_threshold=text_thresh,
                target_sizes=[image_pil.size[::-1]]
            )[0]
        except TypeError:
            return self.processor.post_process_grounded_object_detection(
                outputs, input_ids,
                threshold=box_thresh, text_threshold=text_thresh,
                target_sizes=[image_pil.size[::-1]]
            )[0]

    # ------------------------------------------------------------------
    @staticmethod
    def _parse_prompt(prompt: str) -> list[str]:
        items = []
        for line in prompt.splitlines():
            line = line.strip()
            if not line or line.startswith("#"):
                continue
            for part in line.replace(".", ",").split(","):
                p = part.strip()
                if p:
                    items.append(p)
        seen, unique = set(), []
        for x in items:
            if x not in seen:
                seen.add(x)
                unique.append(x)
        return unique

    # ------------------------------------------------------------------
    @staticmethod
    def _nms(boxes: np.ndarray, scores: np.ndarray, iou_thresh: float) -> list[int]:
        if len(boxes) == 0:
            return []
        x1, y1, x2, y2 = boxes[:,0], boxes[:,1], boxes[:,2], boxes[:,3]
        areas  = (x2 - x1) * (y2 - y1)
        order  = scores.argsort()[::-1]
        keep   = []
        while order.size > 0:
            i = order[0]
            keep.append(int(i))
            xx1 = np.maximum(x1[i], x1[order[1:]])
            yy1 = np.maximum(y1[i], y1[order[1:]])
            xx2 = np.minimum(x2[i], x2[order[1:]])
            yy2 = np.minimum(y2[i], y2[order[1:]])
            w   = np.maximum(0.0, xx2 - xx1)
            h   = np.maximum(0.0, yy2 - yy1)
            iou = (w * h) / (areas[i] + areas[order[1:]] - w * h + 1e-6)
            order = order[np.where(iou <= iou_thresh)[0] + 1]
        return keep


# ──────────────────────────────────────────────────────────
# SAM2Tracker — sliding-window SAM 2 tracking engine
# ──────────────────────────────────────────────────────────
class SAM2Tracker:
    """
    Proper SAM 2 video tracker with:
      1. Sliding-window propagation   — keeps memory bank fresh
      2. Automatic lost-object detection — mask area < MIN_MASK_AREA
      3. DINO re-anchor on lost objects — relocalizes using text prompt
      4. Bbox-only rendering            — fast, VRAM-friendly
    """

    # Palette — visually distinct colors (BGR)
    PALETTE = [
        (255,  80,  80),   # blue-ish
        ( 80, 220,  80),   # green
        ( 80,  80, 255),   # red
        (  0, 220, 220),   # yellow
        (220,   0, 220),   # magenta
        (220, 220,   0),   # cyan
        (255, 160,   0),   # orange
        (160,   0, 200),   # purple
        (  0, 180, 180),   # teal
        (  0, 140, 255),   # gold
        (180, 255,   0),   # lime
        (255,   0, 150),   # pink
    ]

    def __init__(self,
                 sam2_checkpoint: str = SAM2_CKPT_DEFAULT,
                 sam2_cfg: str = SAM2_CFG_DEFAULT,
                 device: Optional[torch.device] = None,
                 chunk_size: int = CHUNK_SIZE_DEFAULT):

        self.device         = device or torch.device(
            "cuda" if torch.cuda.is_available() else "cpu"
        )
        self.sam2_checkpoint = sam2_checkpoint
        self.sam2_cfg        = sam2_cfg
        self.chunk_size      = chunk_size
        self.predictor       = None

    # ------------------------------------------------------------------
    def load(self):
        from sam2.build_sam import build_sam2_video_predictor
        print(f"[SAM2] Loading predictor (device={self.device}) …")
        self.predictor = build_sam2_video_predictor(
            self.sam2_cfg, self.sam2_checkpoint, device=self.device
        )
        print("[SAM2] Loaded.")

    # ------------------------------------------------------------------
    def track_video(self,
                    frame_store:    VideoFrameStore,
                    tracked_objects: list[TrackedObject],
                    dino:           DinoDetector,
                    prompt:         str,
                    box_threshold:  float = 0.30,
                    text_threshold: float = 0.25,
                    iou_threshold:  float = 0.45,
                    output_path:    str = "output.mp4",
                    progress_cb=None) -> list[str]:
        """
        Main entry — runs sliding-window SAM 2 tracking and writes annotated video.

        Returns list of tracked label strings.
        """
        frame_paths = frame_store.frame_paths
        total       = len(frame_paths)
        W, H        = frame_store.width, frame_store.height
        fps         = frame_store.target_fps or frame_store.orig_fps

        if total == 0:
            raise RuntimeError("No frames to track!")
        if not tracked_objects:
            raise RuntimeError("No objects to track — run DINO detection first.")

        # Use browser-friendly H.264 (avc1) codec if possible, fallback to mp4v
        fourcc = cv2.VideoWriter_fourcc(*"avc1")
        writer = cv2.VideoWriter(output_path, fourcc, fps, (W, H))
        if not writer.isOpened():
            print("[WARN] avc1 codec not opened, falling back to mp4v.")
            fourcc = cv2.VideoWriter_fourcc(*"mp4v")
            writer = cv2.VideoWriter(output_path, fourcc, fps, (W, H))

        # ── sliding window loop ──────────────────────────────────────
        # We divide the video into chunks. SAM 2 is initialized fresh
        # at the start of each chunk using the last known box of every object.
        # This keeps the memory bank small and accurate.

        chunk_starts = list(range(0, total, self.chunk_size))
        print(f"\n[Track] {total} frames · {len(chunk_starts)} chunk(s) · "
              f"chunk_size={self.chunk_size}")

        for c_num, chunk_start in enumerate(chunk_starts):
            chunk_end   = min(chunk_start + self.chunk_size, total)
            chunk_paths = frame_paths[chunk_start:chunk_end]
            chunk_len   = len(chunk_paths)

            print(f"\n[Chunk {c_num+1}/{len(chunk_starts)}] "
                  f"frames {chunk_start}–{chunk_end-1} ({chunk_len} frames)")

            # ── 1. Create chunk frames directory using symlinks ──────
            import tempfile, shutil
            chunk_dir = os.path.join(
                os.path.dirname(frame_store.output_dir),
                f"_chunk_{c_num:04d}"
            )
            if os.path.exists(chunk_dir):
                shutil.rmtree(chunk_dir)
            os.makedirs(chunk_dir)
            # Use os.symlink for instantaneous setup and disk space saving
            for local_i, src in enumerate(chunk_paths):
                dst = os.path.join(chunk_dir, f"{local_i:05d}.jpg")
                if not os.path.exists(dst):
                    try:
                        os.symlink(os.path.abspath(src), dst)
                    except Exception:
                        # Fallback to copying if symlink fails
                        shutil.copy2(src, dst)

            # ── 2. Init SAM 2 state for this chunk ──────────────────
            autocast = (torch.autocast("cuda", dtype=torch.bfloat16)
                        if "cuda" in str(self.device) else
                        torch.autocast("cpu",  dtype=torch.float32))

            with torch.inference_mode(), autocast:
                # On macOS (MPS), unified memory makes CPU offloading slow. Disable it.
                offload_video = True
                if "mps" in str(self.device):
                    offload_video = False

                state = self.predictor.init_state(
                    video_path=chunk_dir,
                    offload_video_to_cpu=offload_video,
                    offload_state_to_cpu=False,
                )
                self.predictor.reset_state(state)

                # ── 3. Register all (non-lost) objects at local frame 0 ──
                registered = 0
                for obj in tracked_objects:
                    if obj.lost:
                        print(f"  [SKIP] id={obj.obj_id} '{obj.label}' is lost, "
                              f"will try DINO recovery after this chunk.")
                        continue
                    self.predictor.add_new_points_or_box(
                        inference_state=state,
                        frame_idx=0,              # always local frame 0 of chunk
                        obj_id=obj.obj_id,
                        box=obj.box,
                    )
                    registered += 1
                print(f"  Registered {registered} objects at chunk start.")

                # ── 4. Propagate through chunk ──────────────────────
                # Collect: for each local frame, map obj_id → mask array
                chunk_masks: dict[int, dict[int, np.ndarray]] = {}
                # Also track last seen box per object for carry-forward
                last_box: dict[int, np.ndarray] = {}

                for local_idx, obj_ids, mask_logits in \
                        self.predictor.propagate_in_video(state):

                    frame_masks: dict[int, np.ndarray] = {}
                    for i, obj_id in enumerate(obj_ids):
                        mask = (mask_logits[i] > 0.0).cpu().numpy().squeeze()
                        if mask.ndim == 0:
                            mask = np.zeros((H, W), dtype=bool)
                        frame_masks[int(obj_id)] = mask

                        # Update last known bounding box from mask
                        if mask.sum() >= MIN_MASK_AREA:
                            ys, xs = np.where(mask)
                            new_box = np.array(
                                [xs.min(), ys.min(), xs.max(), ys.max()],
                                dtype=np.float32
                            )
                            last_box[int(obj_id)] = new_box

                    chunk_masks[local_idx] = frame_masks

                    if progress_cb:
                        progress_cb(chunk_start + local_idx + 1, total)

                self.predictor.reset_state(state)

            # ── 5. Update tracked objects with last seen boxes ───────
            # Set to lost if it didn't have a valid mask in the last frame of the chunk
            for obj in tracked_objects:
                last_frame_mask = chunk_masks.get(chunk_len - 1, {}).get(obj.obj_id)
                if last_frame_mask is not None and last_frame_mask.sum() >= MIN_MASK_AREA and obj.obj_id in last_box:
                    obj.box  = last_box[obj.obj_id]
                    obj.lost = False
                    obj.lost_frames = 0
                else:
                    obj.lost = True
                    obj.lost_frames += chunk_len
                    print(f"  [LOST] id={obj.obj_id} '{obj.label}' — not visible at chunk end.")

            # ── 6. DINO recovery for lost objects ────────────────────
            # Run DINO on the LAST frame of this chunk to relocate them
            lost_objects = [o for o in tracked_objects if o.lost]
            if lost_objects:
                last_chunk_frame = chunk_paths[-1]
                print(f"  [Recovery] Running DINO on frame {chunk_end-1} "
                      f"for {len(lost_objects)} lost object(s) …")
                boxes, scores, labels = dino.detect(
                    last_chunk_frame, prompt,
                    box_threshold, text_threshold, iou_threshold
                )
                recovered = self._match_lost_to_dino(
                    lost_objects, boxes, labels, iou_threshold
                )
                for obj_id, new_box in recovered.items():
                    for obj in tracked_objects:
                        if obj.obj_id == obj_id:
                            obj.box  = new_box
                            obj.lost = False
                            obj.lost_frames = 0
                            print(f"  [Recovered] id={obj_id} '{obj.label}' "
                                  f"at chunk boundary.")

            # ── 7. Render and write frames for this chunk ────────────
            for local_idx in range(chunk_len):
                global_idx = chunk_start + local_idx
                frame      = cv2.imread(chunk_paths[local_idx])
                masks_here = chunk_masks.get(local_idx, {})
                
                # --- Save crops of objects before drawing on the frame ---
                for obj in tracked_objects:
                    mask = masks_here.get(obj.obj_id)
                    if mask is not None and mask.sum() >= MIN_MASK_AREA:
                        area = mask.sum()
                        if not hasattr(obj, 'max_mask_area'):
                            obj.max_mask_area = 0
                            obj.best_crop = None
                        if area > obj.max_mask_area:
                            obj.max_mask_area = area
                            ys, xs = np.where(mask)
                            h_f, w_f = frame.shape[:2]
                            bx1, bx2 = max(0, int(xs.min())), min(w_f - 1, int(xs.max()))
                            by1, by2 = max(0, int(ys.min())), min(h_f - 1, int(ys.max()))
                            if bx2 > bx1 and by2 > by1:
                                obj.best_crop = frame[by1:by2+1, bx1:bx2+1].copy()
                
                # Draw annotations on the frame and write it directly to the video file
                frame = self._draw_frame(frame, masks_here, tracked_objects)
                writer.write(frame)

            # cleanup temp chunk dir
            shutil.rmtree(chunk_dir, ignore_errors=True)

            # Clean up memory state and empty PyTorch CUDA/MPS caches to avoid OOM
            if 'state' in locals():
                del state
            import gc
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
            elif hasattr(torch, 'mps') and torch.backends.mps.is_available():
                torch.mps.empty_cache()

            if progress_cb:
                progress_cb(chunk_end, total)

        writer.release()
        print(f"[Done] Saved: {os.path.abspath(output_path)}")
        return [o.label for o in tracked_objects]

    # ------------------------------------------------------------------
    def _draw_frame(self,
                    frame: np.ndarray,
                    masks: dict[int, np.ndarray],
                    tracked_objects: list[TrackedObject]) -> np.ndarray:
        """
        Draw bounding box + label + ID.
        No pixel mask overlay → fast and VRAM-independent.
        """
        if frame is None:
            return frame

        for obj in tracked_objects:
            oid   = obj.obj_id
            color = self.PALETTE[oid % len(self.PALETTE)]
            mask  = masks.get(oid)

            if mask is None or mask.sum() < MIN_MASK_AREA:
                # Object not visible this frame — draw a faded indicator on
                # the last known box location
                x1, y1, x2, y2 = obj.box.astype(int)
                cv2.rectangle(frame, (x1, y1), (x2, y2), color, 1)
                self._put_label(frame, f"{obj.label} #{oid} [?]",
                                x1, y1, color, alpha=0.4)
                continue

            # Derive tight bounding box from the mask
            ys, xs = np.where(mask)
            bx1, bx2 = int(xs.min()), int(xs.max())
            by1, by2 = int(ys.min()), int(ys.max())

            # Draw solid bounding box
            cv2.rectangle(frame, (bx1, by1), (bx2, by2), color, 2)
            self._put_label(frame, f"{obj.label} #{oid}", bx1, by1, color)

        return frame

    # ------------------------------------------------------------------
    @staticmethod
    def _put_label(frame: np.ndarray, text: str,
                   x: int, y: int, color: tuple,
                   alpha: float = 1.0):
        font       = cv2.FONT_HERSHEY_SIMPLEX
        scale      = 0.5
        thickness  = 1
        (tw, th), _ = cv2.getTextSize(text, font, scale, thickness)
        pad        = 4
        bkg_y1     = max(0, y - th - pad * 2)
        bkg_y2     = y
        bkg_x2     = x + tw + pad * 2

        # Background rectangle
        if alpha >= 0.9:
            cv2.rectangle(frame, (x, bkg_y1), (bkg_x2, bkg_y2), color, -1)
        else:
            overlay = frame.copy()
            cv2.rectangle(overlay, (x, bkg_y1), (bkg_x2, bkg_y2), color, -1)
            cv2.addWeighted(overlay, alpha, frame, 1 - alpha, 0, frame)

        # Text
        cv2.putText(frame, text, (x + pad, y - pad),
                    font, scale, (255, 255, 255), thickness, cv2.LINE_AA)

    # ------------------------------------------------------------------
    @staticmethod
    def _match_lost_to_dino(lost_objects: list[TrackedObject],
                             dino_boxes:   np.ndarray,
                             dino_labels:  list[str],
                             iou_threshold: float = 0.20
                             ) -> dict[int, np.ndarray]:
        """
        For each lost object, find the best DINO detection that:
          (a) has the same label (or close substring match), AND
          (b) overlaps reasonably OR is the closest available detection.
        Returns {obj_id: new_box}.
        """
        recovered = {}
        used_dino = set()

        for obj in lost_objects:
            best_idx   = None
            best_score = -1.0

            for d_idx, (d_box, d_label) in enumerate(zip(dino_boxes, dino_labels)):
                if d_idx in used_dino:
                    continue

                # Label similarity: simple substring match
                label_ok = (obj.label.lower() in d_label.lower() or
                            d_label.lower() in obj.label.lower())
                if not label_ok:
                    continue

                # Prefer boxes overlapping the last known location
                x1, y1, x2, y2   = obj.box
                dx1,dy1,dx2,dy2  = d_box
                ix1 = max(x1, dx1); iy1 = max(y1, dy1)
                ix2 = min(x2, dx2); iy2 = min(y2, dy2)
                iw  = max(0, ix2 - ix1); ih  = max(0, iy2 - iy1)
                inter = iw * ih
                union = ((x2-x1)*(y2-y1) + (dx2-dx1)*(dy2-dy1) - inter + 1e-6)
                iou   = inter / union

                # Score: label match + IoU bonus
                score = 0.5 + iou
                if score > best_score:
                    best_score = score
                    best_idx   = d_idx

            if best_idx is not None:
                recovered[obj.obj_id] = dino_boxes[best_idx].astype(np.float32)
                used_dino.add(best_idx)

        return recovered