scorevision: push artifact

miner.py

@@ -1,22 +1,4 @@
-"""TurboVision crime-detection miner.
-
-YOLO11s @ 1280x1280, 6-class detection (balaclava, bat, glove, graffiti, hoodie,
-spray paint), ONNX with end-to-end NMS baked in.
-
-Output of weights.onnx: [1, 300, 6] = x1, y1, x2, y2, conf, cls (post-NMS).
-
-Inference pipeline:
-  1) Primary forward pass on the full image.
-  2) Hflip TTA: forward on horizontally-flipped image, transform boxes back.
-  3) Per-class hard-NMS to merge primary + flip outputs.
-  4) Cross-class IoU dedup (suppresses same physical object getting two class labels).
-  5) Consensus-confidence boost: when both views agree on a cluster, take max score.
-  6) Sanity filter (min size, aspect ratio).
-
-Class taxonomy (must match the validator manifest's `objects` list for this element):
-  0 balaclava   1 bat   2 glove   3 graffiti   4 hoodie   5 spray paint
-"""
-
+"""ScoreVision crime detector — YOLOv11s with flip TTA + per-class conf + cross-class NMS."""
 from pathlib import Path
 import math
 
@@ -43,43 +25,40 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
+    """ONNX Runtime miner with horizontal-flip TTA and per-class confidence."""
+
+    class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
+    input_size = 1280
+    iou_thres = 0.45
+    cross_iou_thresh = 0.50
+    max_aspect_ratio = 10.0
+    max_det = 150
+    # tuned on held-out SAM3-labeled crime set
+    _conf_thres_array = np.array(
+        [0.50, 0.50, 0.30, 0.30, 0.40, 0.40], dtype=np.float32
+    )
+
     def __init__(self, path_hf_repo: Path) -> None:
         model_path = path_hf_repo / "weights.onnx"
-
-        # Validator manifest order (from spec.json `objects`):
-        #   0=balaclava 1=hoodie 2=glove 3=bat 4="spray paint" 5=graffiti
-        # v5 weights.onnx was trained with this exact order, so cls_remap is identity.
-        cn_path = model_path.with_name("class_names.txt")
-        if cn_path.is_file():
-            self.class_names = [
-                ln.strip()
-                for ln in cn_path.read_text(encoding="utf-8").splitlines()
-                if ln.strip() and not ln.strip().startswith("#")
-            ]
-        else:
-            self.class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
-        self.cls_remap = np.arange(len(self.class_names), dtype=np.int32)
-
         print("ORT version:", ort.__version__)
         try:
             ort.preload_dlls()
-            print("✅ onnxruntime.preload_dlls() success")
+            print("preload_dlls success")
         except Exception as e:
-            print(f"⚠️ preload_dlls failed: {e}")
-        print("ORT available providers BEFORE session:", ort.get_available_providers())
+            print(f"preload_dlls failed: {e}")
+        print("ORT available providers:", ort.get_available_providers())
 
         sess_options = ort.SessionOptions()
         sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-
         try:
             self.session = ort.InferenceSession(
                 str(model_path),
                 sess_options=sess_options,
                 providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
             )
-            print("✅ Created ORT session with preferred CUDA provider list")
+            print("Created ORT session with preferred CUDA provider list")
         except Exception as e:
-            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
+            print(f"CUDA session creation failed, falling back to CPU: {e}")
             self.session = ort.InferenceSession(
                 str(model_path),
                 sess_options=sess_options,
@@ -87,73 +66,46 @@ class Miner:
             )
         print("ORT session providers:", self.session.get_providers())
 
-        inp = self.session.get_inputs()[0]
-        self.input_name = inp.name
-        self.output_names = [o.name for o in self.session.get_outputs()]
-        self.input_shape = inp.shape
-        self.input_dtype = np.float16 if "float16" in inp.type else np.float32
-
-        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
-        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+        for inp in self.session.get_inputs():
+            print("INPUT:", inp.name, inp.shape, inp.type)
+        for out in self.session.get_outputs():
+            print("OUTPUT:", out.name, out.shape, out.type)
 
-        # Tuning matched to alfred's deployed model — bias toward precision to dodge
-        # the false_positive pillar penalty (validator weights FP heavily on this element).
-        # v13 sweet spot on starter (true GT): uniform conf=0.50.
-        # Tuning per-class on 7 images overfits — leave-one-out CV showed it
-        # collapsed to 0.314 on held-out shards. Uniform 0.50 is robust.
-        self.conf_thres = 0.50
-        self.conf_thres_per_class = np.array([0.50] * 6, dtype=np.float32)
-        self.iou_thres = 0.4
-        self.cross_iou_thresh = 0.7
-        self.max_det = 100
-        self.use_tta = False
-
-        # Sanity filter — reject obviously bad boxes
-        self.min_box_area = 14 * 14
-        self.min_side = 8
-        self.max_aspect_ratio = 8.0
-        self.max_box_area_ratio = 0.95
-
-        print(f"✅ ONNX loaded: {model_path}")
-        print(f"✅ providers: {self.session.get_providers()}")
-        print(f"✅ input: name={self.input_name}, shape={self.input_shape}, dtype={self.input_dtype}")
-        print(f"✅ classes: {self.class_names}")
-        print(f"✅ config: conf={self.conf_thres}, iou={self.iou_thres}, "
-              f"cross_iou={self.cross_iou_thresh}, TTA={self.use_tta}")
+        self.input_name = self.session.get_inputs()[0].name
+        self.output_names = [o.name for o in self.session.get_outputs()]
+        self.input_shape = self.session.get_inputs()[0].shape
+        self.input_height = self._safe_dim(self.input_shape[2], self.input_size)
+        self.input_width = self._safe_dim(self.input_shape[3], self.input_size)
+        print(f"ONNX loaded: {model_path}  input={self.input_shape}")
+        print(
+            "per-class conf: "
+            + ", ".join(
+                f"{n}={t:.3f}" for n, t in zip(self.class_names, self._conf_thres_array.tolist())
+            )
+        )
 
     def __repr__(self) -> str:
-        return (
-            f"ONNXRuntime(session={type(self.session).__name__}, "
-            f"providers={self.session.get_providers()})"
-        )
+        return f"ONNXRuntime(providers={self.session.get_providers()})"
 
     @staticmethod
     def _safe_dim(value, default: int) -> int:
         return value if isinstance(value, int) and value > 0 else default
 
-    def _letterbox(
-        self,
-        image: ndarray,
-        new_shape: tuple[int, int],
-        color=(114, 114, 114),
-    ) -> tuple[ndarray, float, tuple[float, float]]:
+    def _letterbox(self, image: ndarray, new_shape: tuple[int, int],
+                   color=(114, 114, 114)
+                   ) -> tuple[ndarray, float, tuple[float, float]]:
         h, w = image.shape[:2]
         new_w, new_h = new_shape
         ratio = min(new_w / w, new_h / h)
-        resized_w = int(round(w * ratio))
-        resized_h = int(round(h * ratio))
-        if (resized_w, resized_h) != (w, h):
+        rw, rh = int(round(w * ratio)), int(round(h * ratio))
+        if (rw, rh) != (w, h):
             interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
-            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)
-        dw = (new_w - resized_w) / 2.0
-        dh = (new_h - resized_h) / 2.0
-        left = int(round(dw - 0.1))
-        right = int(round(dw + 0.1))
-        top = int(round(dh - 0.1))
-        bottom = int(round(dh + 0.1))
+            image = cv2.resize(image, (rw, rh), interpolation=interp)
+        dw, dh = (new_w - rw) / 2.0, (new_h - rh) / 2.0
+        top, bot = int(round(dh - 0.1)), int(round(dh + 0.1))
+        lt, rt = int(round(dw - 0.1)), int(round(dw + 0.1))
         padded = cv2.copyMakeBorder(
-            image, top, bottom, left, right,
-            borderType=cv2.BORDER_CONSTANT, value=color,
+            image, top, bot, lt, rt, cv2.BORDER_CONSTANT, value=color
         )
         return padded, ratio, (dw, dh)
 
@@ -161,445 +113,195 @@ class Miner:
         orig_h, orig_w = image.shape[:2]
         img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height))
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        img = img.astype(self.input_dtype) / 255.0
+        img = img.astype(np.float32) / 255.0
         img = np.transpose(img, (2, 0, 1))[None, ...]
-        img = np.ascontiguousarray(img)
+        img = np.ascontiguousarray(img, dtype=np.float32)
         return img, ratio, pad, (orig_w, orig_h)
 
     @staticmethod
-    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
-        w, h = image_size
+    def _clip(boxes, size):
+        w, h = size
         boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)
         boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)
         boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)
         boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)
         return boxes
 
-    def _filter_sane_boxes(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        orig_size: tuple[int, int],
-    ):
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
-        orig_w, orig_h = orig_size
-        image_area = float(orig_w * orig_h)
+    @staticmethod
+    def _hard_nms(boxes, scores, iou_thr):
+        n = len(boxes)
+        if n == 0:
+            return np.array([], dtype=np.intp)
+        order = np.argsort(-scores)
         keep = []
-        for i, box in enumerate(boxes):
-            x1, y1, x2, y2 = box.tolist()
-            bw = x2 - x1
-            bh = y2 - y1
-            if bw <= 0 or bh <= 0:
-                continue
-            if bw < self.min_side or bh < self.min_side:
-                continue
-            area = bw * bh
-            if area < self.min_box_area:
-                continue
-            if area > self.max_box_area_ratio * image_area:
-                continue
-            ar = max(bw / max(bh, 1e-6), bh / max(bw, 1e-6))
-            if ar > self.max_aspect_ratio:
-                continue
+        while len(order) > 0:
+            i = int(order[0])
             keep.append(i)
-        if not keep:
-            return (
-                np.empty((0, 4), dtype=np.float32),
-                np.empty((0,), dtype=np.float32),
-                np.empty((0,), dtype=np.int32),
+            if len(order) == 1:
+                break
+            rest = order[1:]
+            xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])
+            yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])
+            xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])
+            yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = max(0.0, boxes[i, 2] - boxes[i, 0]) * max(0.0, boxes[i, 3] - boxes[i, 1])
+            a_r = (
+                np.maximum(0.0, boxes[rest, 2] - boxes[rest, 0])
+                * np.maximum(0.0, boxes[rest, 3] - boxes[rest, 1])
             )
-        k = np.array(keep, dtype=np.intp)
-        return boxes[k], scores[k], cls_ids[k]
-
-    @staticmethod
-    def _hard_nms(
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        N = len(boxes)
-        if N == 0:
-            return np.array([], dtype=np.intp)
-        boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        order = np.argsort(scores)[::-1]
-        keep: list[int] = []
-        suppressed = np.zeros(N, dtype=bool)
-        for i in range(N):
-            idx = order[i]
-            if suppressed[idx]:
-                continue
-            keep.append(int(idx))
-            bi = boxes[idx]
-            for k in range(i + 1, N):
-                jdx = order[k]
-                if suppressed[jdx]:
-                    continue
-                bj = boxes[jdx]
-                xx1 = max(bi[0], bj[0])
-                yy1 = max(bi[1], bj[1])
-                xx2 = min(bi[2], bj[2])
-                yy2 = min(bi[3], bj[3])
-                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
-                area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-                area_j = (bj[2] - bj[0]) * (bj[3] - bj[1])
-                iou = inter / (area_i + area_j - inter + 1e-7)
-                if iou > iou_thresh:
-                    suppressed[jdx] = True
+            iou = inter / (a_i + a_r - inter + 1e-7)
+            order = rest[iou <= iou_thr]
         return np.array(keep, dtype=np.intp)
 
-    def _per_class_hard_nms(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
+    def _per_class_hard_nms(self, boxes, scores, cls_ids, iou_thr):
         if len(boxes) == 0:
             return np.array([], dtype=np.intp)
-        all_keep: list[int] = []
+        keep_all = []
         for c in np.unique(cls_ids):
             mask = cls_ids == c
-            indices = np.where(mask)[0]
-            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)
-            all_keep.extend(indices[keep].tolist())
-        all_keep.sort()
-        return np.array(all_keep, dtype=np.intp)
-
-    @staticmethod
-    def _cross_class_dedup(
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        iou_thresh: float,
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+            idx = np.where(mask)[0]
+            keep = self._hard_nms(boxes[mask], scores[mask], iou_thr)
+            keep_all.extend(idx[keep].tolist())
+        keep_all.sort()
+        return np.array(keep_all, dtype=np.intp)
+
+    def _cross_class_dedup(self, boxes, scores, cls_ids, iou_thr):
+        """If two boxes (any class) heavily overlap, keep only the higher-scoring one."""
         n = len(boxes)
-        if n <= 1:
-            return boxes, scores, cls_ids
-        boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        cls_ids = np.asarray(cls_ids, dtype=np.int32)
-        areas = np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) * np.maximum(
-            0.0, boxes[:, 3] - boxes[:, 1]
-        )
-        # Keep larger boxes first, then higher score.
-        order = np.lexsort((-scores, -areas))
+        if n == 0:
+            return np.array([], dtype=np.intp)
+        order = np.argsort(-scores)
+        keep = []
         suppressed = np.zeros(n, dtype=bool)
-        keep: list[int] = []
         for i in order:
             if suppressed[i]:
                 continue
             keep.append(int(i))
-            bi = boxes[i]
-            xx1 = np.maximum(bi[0], boxes[:, 0])
-            yy1 = np.maximum(bi[1], boxes[:, 1])
-            xx2 = np.minimum(bi[2], boxes[:, 2])
-            yy2 = np.minimum(bi[3], boxes[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            area_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
-            union = area_i + areas - inter + 1e-7
-            iou = inter / union
-            dup = iou > iou_thresh
-            dup[i] = False
-            suppressed |= dup
-        keep_idx = np.array(keep, dtype=np.intp)
-        return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]
-
-    @staticmethod
-    def _max_score_per_cluster(
-        coords: np.ndarray,
-        scores: np.ndarray,
-        keep_indices: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        n_keep = len(keep_indices)
-        if n_keep == 0:
-            return np.array([], dtype=np.float32)
-        coords = np.asarray(coords, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        out = np.empty(n_keep, dtype=np.float32)
-        for i in range(n_keep):
-            idx = keep_indices[i]
-            bi = coords[idx]
-            xx1 = np.maximum(bi[0], coords[:, 0])
-            yy1 = np.maximum(bi[1], coords[:, 1])
-            xx2 = np.minimum(bi[2], coords[:, 2])
-            yy2 = np.minimum(bi[3], coords[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-            areas_j = (coords[:, 2] - coords[:, 0]) * (coords[:, 3] - coords[:, 1])
-            iou = inter / (area_i + areas_j - inter + 1e-7)
-            in_cluster = iou >= iou_thresh
-            out[i] = float(np.max(scores[in_cluster]))
-        return out
+            ix1 = np.maximum(boxes[i, 0], boxes[:, 0])
+            iy1 = np.maximum(boxes[i, 1], boxes[:, 1])
+            ix2 = np.minimum(boxes[i, 2], boxes[:, 2])
+            iy2 = np.minimum(boxes[i, 3], boxes[:, 3])
+            inter = np.maximum(0.0, ix2 - ix1) * np.maximum(0.0, iy2 - iy1)
+            a_i = max(0.0, boxes[i, 2] - boxes[i, 0]) * max(0.0, boxes[i, 3] - boxes[i, 1])
+            a_r = (
+                np.maximum(0.0, boxes[:, 2] - boxes[:, 0])
+                * np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
+            )
+            iou = inter / (a_i + a_r - inter + 1e-7)
+            iou[i] = 0.0
+            suppressed |= iou >= iou_thr
+        return np.array(keep, dtype=np.intp)
 
-    def _decode_raw_dets(
-        self,
-        preds: np.ndarray,
-        ratio: float,
-        pad: tuple[float, float],
-        orig_size: tuple[int, int],
-        *,
-        apply_conf_thresh: bool = True,
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Decode end2end NMS output and return (boxes, scores, cls_ids)
-        in original image coordinates, after conf-threshold + remap + letterbox-reverse + sanity.
+    def _filter_sane(self, boxes, scores, cls_ids, orig_size):
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+        w, h = orig_size
+        area_img = float(w * h)
+        bw = np.maximum(0.0, boxes[:, 2] - boxes[:, 0])
+        bh = np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
+        area = bw * bh
+        ar = np.where(
+            (bw > 0) & (bh > 0),
+            np.maximum(bw / np.maximum(bh, 1e-6), bh / np.maximum(bw, 1e-6)),
+            np.inf,
+        )
+        keep = (area <= 0.95 * area_img) & (ar <= self.max_aspect_ratio)
+        return boxes[keep], scores[keep], cls_ids[keep]
 
-        When apply_conf_thresh=False, the conf-threshold filter is skipped (used for
-        the no-detection fallback path: take the single top-conf raw box)."""
+    def _decode(self, preds: ndarray, ratio: float, pad: tuple[float, float],
+                orig_size: tuple[int, int]) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """ONNX output is [1,300,6]: x1,y1,x2,y2,conf,cls in letterboxed coords."""
         if preds.ndim == 3 and preds.shape[0] == 1:
             preds = preds[0]
         if preds.ndim != 2 or preds.shape[1] < 6:
-            raise ValueError(f"Unexpected ONNX output shape: {preds.shape}")
-
-        boxes = preds[:, :4].astype(np.float32)
-        scores = preds[:, 4].astype(np.float32)
-        cls_ids = preds[:, 5].astype(np.int32)
-
-        valid = (cls_ids >= 0) & (cls_ids < len(self.cls_remap)) & (scores > 0)
-        boxes, scores, cls_ids = boxes[valid], scores[valid], cls_ids[valid]
-        cls_ids = self.cls_remap[cls_ids]
-
-        if apply_conf_thresh:
-            # Per-class threshold: each box compared against its own class's threshold
-            cls_thresh = np.full(len(scores), self.conf_thres, dtype=np.float32)
-            valid_cls = (cls_ids >= 0) & (cls_ids < len(self.conf_thres_per_class))
-            cls_thresh[valid_cls] = self.conf_thres_per_class[cls_ids[valid_cls]]
-            keep = scores >= cls_thresh
-            boxes = boxes[keep]
-            scores = scores[keep]
-            cls_ids = cls_ids[keep]
-        if len(boxes) == 0:
             return (
                 np.empty((0, 4), dtype=np.float32),
                 np.empty((0,), dtype=np.float32),
                 np.empty((0,), dtype=np.int32),
             )
-
+        boxes = preds[:, :4].astype(np.float32)
+        scores = preds[:, 4].astype(np.float32)
+        cls_ids = preds[:, 5].astype(np.int32)
+        # drop padded rows
+        keep = (scores > 0) & (cls_ids >= 0) & (cls_ids < len(self.class_names))
+        boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+        # per-class confidence
+        thr = self._conf_thres_array[cls_ids]
+        keep = scores >= thr
+        boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+        # untransform: subtract pad, divide ratio
         pad_w, pad_h = pad
-        orig_w, orig_h = orig_size
         boxes[:, [0, 2]] -= pad_w
         boxes[:, [1, 3]] -= pad_h
         boxes /= ratio
-        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
-
-        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
+        boxes = self._clip(boxes, orig_size)
         return boxes, scores, cls_ids
 
-    def _forward(
-        self, image: np.ndarray
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    def _predict_single(self, image: ndarray):
         x, ratio, pad, orig_size = self._preprocess(image)
         out = self.session.run(self.output_names, {self.input_name: x})[0]
-        return self._decode_raw_dets(out, ratio, pad, orig_size)
+        return self._decode(out, ratio, pad, orig_size)
 
-    def _forward_with_fallback(
-        self, image: np.ndarray
-    ) -> tuple[
-        tuple[np.ndarray, np.ndarray, np.ndarray],
-        tuple[np.ndarray, np.ndarray, np.ndarray],
-    ]:
-        """Run ONNX once, decode twice: (filtered @ conf_thres, all-survived sanity)."""
-        x, ratio, pad, orig_size = self._preprocess(image)
-        out = self.session.run(self.output_names, {self.input_name: x})[0]
-        primary = self._decode_raw_dets(out, ratio, pad, orig_size, apply_conf_thresh=True)
-        fallback = self._decode_raw_dets(out, ratio, pad, orig_size, apply_conf_thresh=False)
-        return primary, fallback
-
-    def _predict_single(self, image: np.ndarray) -> list[BoundingBox]:
-        (boxes, scores, cls_ids), (fb_b, fb_s, fb_c) = self._forward_with_fallback(image)
-        ih, iw = image.shape[:2]
-        if len(boxes) > 0:
-            return self._build_results(boxes, scores, cls_ids, image_size=(iw, ih))
-        # FALLBACK: nothing passed conf_thres — return single top-conf box
-        # (any class, any conf > 0) so the validator's mAP isn't a hard zero.
-        if len(fb_b) == 0:
+    def _predict_tta(self, image: ndarray) -> list[BoundingBox]:
+        b0, s0, c0 = self._predict_single(image)
+        flipped = cv2.flip(image, 1)
+        bf, sf, cf = self._predict_single(flipped)
+        if len(bf):
+            w = image.shape[1]
+            bf2 = bf.copy()
+            bf2[:, 0] = w - bf[:, 2]
+            bf2[:, 2] = w - bf[:, 0]
+            bf = bf2
+        boxes = np.concatenate([b0, bf], axis=0) if len(b0) or len(bf) else b0
+        scores = np.concatenate([s0, sf], axis=0) if len(b0) or len(bf) else s0
+        cls_ids = np.concatenate([c0, cf], axis=0) if len(b0) or len(bf) else c0
+        if len(boxes) == 0:
             return []
-        i = int(np.argmax(fb_s))
-        return self._build_results(
-            fb_b[i:i + 1], fb_s[i:i + 1], fb_c[i:i + 1], image_size=(iw, ih)
+        # filter + per-class NMS + cross-class dedup
+        boxes, scores, cls_ids = self._filter_sane(
+            boxes, scores, cls_ids, (image.shape[1], image.shape[0])
         )
-
-    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
-        """Hflip TTA: merge primary + flipped via per-class hard-NMS,
-        then cross-class dedup, with consensus-confidence boost."""
-        ow = image.shape[1]
-        b1, s1, c1 = self._forward(image)
-
-        flipped = cv2.flip(image, 1)
-        b2, s2, c2 = self._forward(flipped)
-        if len(b2):
-            x1f = ow - b2[:, 2]
-            x2f = ow - b2[:, 0]
-            b2 = np.stack([x1f, b2[:, 1], x2f, b2[:, 3]], axis=1)
-
-        if len(b1) == 0 and len(b2) == 0:
+        if len(boxes) == 0:
             return []
-
-        boxes = np.concatenate([b1, b2], axis=0) if len(b2) else b1
-        scores = np.concatenate([s1, s2], axis=0) if len(b2) else s1
-        cls_ids = np.concatenate([c1, c2], axis=0) if len(b2) else c1
-
         keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
         if len(keep) == 0:
             return []
-        keep = keep[: self.max_det]
-
-        # Consensus-confidence boost: cluster by IoU and take max score.
-        boosted = self._max_score_per_cluster(boxes, scores, keep, self.iou_thres)
-
-        boxes = boxes[keep]
-        cls_ids = cls_ids[keep]
-        scores = boosted
-
-        boxes, scores, cls_ids = self._cross_class_dedup(
-            boxes, scores, cls_ids, self.cross_iou_thresh
-        )
-        if len(boxes) == 0:
-            return []
-
-        ih, iw = image.shape[:2]
-        return self._build_results(boxes, scores, cls_ids, image_size=(iw, ih))
-
-    def _filter_balaclava_geometry(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        image_size: tuple[int, int] | None = None,
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        # Real-balaclava prior (from 43 manual GT labels):
-        #   aspect ratio max(w/h, h/w):  p5=1.11, median=1.33, p99=1.71
-        #   rel area % of image:         p1=0.041, p5=0.070, p10=0.087
-        # FP balaclavas frequently violate these (very thin/wide boxes from
-        # face-fragment matches, or tiny ~0.01%-area boxes from texture noise).
-        BALACLAVA = 0
-        ASPECT_MAX = 1.8     # above p99 of real
-        REL_AREA_MIN = 0.0004  # below p1 of real (0.04%)
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
-        is_bal = cls_ids == BALACLAVA
-        if not is_bal.any():
-            return boxes, scores, cls_ids
-        keep = np.ones(len(boxes), dtype=bool)
-        if image_size is not None:
-            iw, ih = image_size
-            img_area = max(1.0, iw * ih)
-        else:
-            img_area = None
-        for i in np.where(is_bal)[0]:
-            x1, y1, x2, y2 = boxes[i]
-            bw = max(1.0, x2 - x1)
-            bh = max(1.0, y2 - y1)
-            aspect = max(bw / bh, bh / bw)
-            if aspect > ASPECT_MAX:
-                keep[i] = False
-                continue
-            if img_area is not None:
-                rel = (bw * bh) / img_area
-                if rel < REL_AREA_MIN:
-                    keep[i] = False
-        return boxes[keep], scores[keep], cls_ids[keep]
-
-    def _suppress_balaclava_under_hoodie(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        # Validator rule: "balaclavas worn under a hoodie hood are IGNORED
-        # (a hoodie includes the jacket and its hood)". A small balaclava
-        # box can sit fully inside a much larger hoodie box — IoU between
-        # them stays low (intersection / large union), but containment
-        # (intersection / balaclava_area) is ~1.0. So drop any balaclava
-        # whose containment by any hoodie box is >= COVER_THRESH.
-        BALACLAVA, HOODIE = 0, 1
-        COVER_THRESH = 0.5
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
-        is_hood = cls_ids == HOODIE
-        is_bal  = cls_ids == BALACLAVA
-        if not is_hood.any() or not is_bal.any():
-            return boxes, scores, cls_ids
-        hood_boxes = boxes[is_hood]
-        keep = np.ones(len(boxes), dtype=bool)
-        for i in np.where(is_bal)[0]:
-            bx1, by1, bx2, by2 = boxes[i]
-            bal_area = max(1.0, (bx2 - bx1) * (by2 - by1))
-            ix1 = np.maximum(bx1, hood_boxes[:, 0])
-            iy1 = np.maximum(by1, hood_boxes[:, 1])
-            ix2 = np.minimum(bx2, hood_boxes[:, 2])
-            iy2 = np.minimum(by2, hood_boxes[:, 3])
-            iw = np.clip(ix2 - ix1, 0.0, None)
-            ih = np.clip(iy2 - iy1, 0.0, None)
-            inter = iw * ih
-            cover = inter / bal_area
-            if (cover >= COVER_THRESH).any():
-                keep[i] = False
-        return boxes[keep], scores[keep], cls_ids[keep]
-
-    def _build_results(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        image_size: tuple[int, int] | None = None,
-    ) -> list[BoundingBox]:
-        boxes, scores, cls_ids = self._filter_balaclava_geometry(
-            boxes, scores, cls_ids, image_size
-        )
-        boxes, scores, cls_ids = self._suppress_balaclava_under_hoodie(
-            boxes, scores, cls_ids
-        )
-        results: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, cls_ids):
-            x1, y1, x2, y2 = box.tolist()
-            if x2 <= x1 or y2 <= y1:
-                continue
-            results.append(
-                BoundingBox(
-                    x1=int(math.floor(x1)),
-                    y1=int(math.floor(y1)),
-                    x2=int(math.ceil(x2)),
-                    y2=int(math.ceil(y2)),
-                    cls_id=int(cls_id),
-                    conf=float(conf),
-                )
+        boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        keep = self._cross_class_dedup(boxes, scores, cls_ids, self.cross_iou_thresh)
+        boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+        if len(scores) > self.max_det:
+            top = np.argsort(-scores)[: self.max_det]
+            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
+        return [
+            BoundingBox(
+                x1=int(math.floor(b[0])),
+                y1=int(math.floor(b[1])),
+                x2=int(math.ceil(b[2])),
+                y2=int(math.ceil(b[3])),
+                cls_id=int(c),
+                conf=float(s),
             )
-        return results
+            for b, s, c in zip(boxes, scores, cls_ids)
+            if b[2] > b[0] and b[3] > b[1]
+        ]
 
-    def predict_batch(
-        self,
-        batch_images: list[ndarray],
-        offset: int,
-        n_keypoints: int,
-    ) -> list[TVFrameResult]:
+    def predict_batch(self, batch_images: list[ndarray], offset: int,
+                      n_keypoints: int) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
-        for frame_number_in_batch, image in enumerate(batch_images):
-            if image is None or not isinstance(image, np.ndarray) or image.ndim != 3:
-                results.append(
-                    TVFrameResult(
-                        frame_id=offset + frame_number_in_batch,
-                        boxes=[],
-                        keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
-                    )
-                )
-                continue
-            if image.dtype != np.uint8:
-                image = image.astype(np.uint8)
+        for j, image in enumerate(batch_images):
             try:
-                if self.use_tta:
-                    boxes = self._predict_tta(image)
-                else:
-                    boxes = self._predict_single(image)
+                boxes = self._predict_tta(image)
             except Exception as e:
-                print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
+                print(f"Inference failed for frame {offset + j}: {e}")
                 boxes = []
             results.append(
                 TVFrameResult(
-                    frame_id=offset + frame_number_in_batch,
+                    frame_id=offset + j,
                     boxes=boxes,
                     keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
                 )