dfine_jina_pipeline.py

""" Pipeline: D-FINE (person/car only) → group detections → crop regions →
find all bboxes inside each crop → Jina-CLIP-v2 embeddings and classification.
Outputs jina_crops folder and results CSV.
"""

import argparse
import csv
import time
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
from transformers import AutoImageProcessor, DFineForObjectDetection

# Jina-CLIP-v2 few-shot (same refs + classify as jina_fewshot.py)
from jina_fewshot import (
    IMAGE_EXTS,
    TRUNCATE_DIM,
    JinaCLIPv2Encoder,
    build_refs,
    classify as jina_classify,
    draw_bboxes_on_image,
    draw_label_on_image,
)

# Only these ref classes get bboxes on group crops and appear in the known-object gallery
KNOWN_DISPLAY_CLASSES = {"gun", "knife", "cigarette", "phone"}
# Only show objects (and group crops) with confidence >= this
MIN_DISPLAY_CONF = 0.7
# Person/car detections must have confidence > this to be used for grouping
PERSON_CAR_MIN_CONF = 0.9

# -----------------------------------------------------------------------------
# Detection + grouping (from reference_detection.py)
# -----------------------------------------------------------------------------

def get_box_dist(box1, box2):
    """Euclidean distance between box centers. box = [x1, y1, x2, y2]."""
    c1 = np.array([(box1[0] + box1[2]) / 2, (box1[1] + box1[3]) / 2])
    c2 = np.array([(box2[0] + box2[2]) / 2, (box2[1] + box2[3]) / 2])
    return np.linalg.norm(c1 - c2)


def group_detections(detections, threshold):
    """
    Group detections by proximity (center distance < threshold).

    detections: list of {"box": [x1,y1,x2,y2], "conf", "cls", ...}
    Returns list of {"box": merged [x1,y1,x2,y2], "conf": best in group, "cls": best in group}.
    """
    if not detections:
        return []

    boxes = [d["box"] for d in detections]
    n = len(boxes)
    adj = {i: [] for i in range(n)}

    for i in range(n):
        for j in range(i + 1, n):
            if get_box_dist(boxes[i], boxes[j]) < threshold:
                adj[i].append(j)
                adj[j].append(i)

    groups = []
    visited = [False] * n

    for i in range(n):
        if not visited[i]:
            group_indices = []
            stack = [i]
            visited[i] = True

            while stack:
                curr = stack.pop()
                group_indices.append(curr)

                for neighbor in adj[curr]:
                    if not visited[neighbor]:
                        visited[neighbor] = True
                        stack.append(neighbor)

            group_dets = [detections[k] for k in group_indices]
            x1 = min(d["box"][0] for d in group_dets)
            y1 = min(d["box"][1] for d in group_dets)
            x2 = max(d["box"][2] for d in group_dets)
            y2 = max(d["box"][3] for d in group_dets)

            best_det = max(group_dets, key=lambda x: x["conf"])
            groups.append({
                "box": [x1, y1, x2, y2],
                "conf": best_det["conf"],
                "cls": best_det["cls"],
                "label": best_det.get("label", str(best_det["cls"])),
            })

    return groups


def box_center_inside(box, crop_box):
    """True if center of box is inside crop_box. All [x1,y1,x2,y2]."""
    cx = (box[0] + box[2]) / 2
    cy = (box[1] + box[3]) / 2
    return (
        crop_box[0] <= cx <= crop_box[2]
        and crop_box[1] <= cy <= crop_box[3]
    )


def expand_box_by_margin(box, margin_ratio, img_w, img_h):
    """Expand box [x1,y1,x2,y2] by margin_ratio (e.g. 0.1 = 10%) on all sides, clamped to image."""
    x1, y1, x2, y2 = box
    w, h = x2 - x1, y2 - y1
    if w <= 0 or h <= 0:
        return box
    mx = w * margin_ratio
    my = h * margin_ratio
    x1 = max(0, x1 - mx)
    y1 = max(0, y1 - my)
    x2 = min(img_w, x2 + mx)
    y2 = min(img_h, y2 + my)
    return [x1, y1, x2, y2]


# 10% margin on person/car group crop (expand crop before running D-FINE on it)
PERSON_CAR_GROUP_MARGIN = 0.10
# Min side (px) for object crops extracted from person/car crop before sending to classifier (objects in crop are larger)
MIN_OBJECT_CROP_SIDE = 112


def squarify_crop_box(bx1, by1, bx2, by2, img_w, img_h):
    """
    Expand the shorter side to match the longer (same ratio / square), centered, clamped to image.
    If height > width: expand width. If width >= height: expand height.
    Returns (bx1, by1, bx2, by2) as integers.
    """
    orig = (int(bx1), int(by1), int(bx2), int(by2))
    w = bx2 - bx1
    h = by2 - by1

    if w <= 0 or h <= 0:
        return orig

    if h > w:
        add = (h - w) / 2.0
        bx1 = max(0, bx1 - add)
        bx2 = min(img_w, bx2 + add)
    else:
        add = (w - h) / 2.0
        by1 = max(0, by1 - add)
        by2 = min(img_h, by2 + add)

    bx1, by1, bx2, by2 = int(bx1), int(by1), int(bx2), int(by2)

    if bx2 <= bx1 or by2 <= by1:
        return orig

    return bx1, by1, bx2, by2


def box_iou(box1, box2):
    """IoU of two boxes [x1,y1,x2,y2]. Returns float in [0, 1]."""
    ix1 = max(box1[0], box2[0])
    iy1 = max(box1[1], box2[1])
    ix2 = min(box1[2], box2[2])
    iy2 = min(box1[3], box2[3])

    inter_w = max(0, ix2 - ix1)
    inter_h = max(0, iy2 - iy1)
    inter = inter_w * inter_h

    a1 = (box1[2] - box1[0]) * (box1[3] - box1[1])
    a2 = (box2[2] - box2[0]) * (box2[3] - box2[1])
    union = a1 + a2 - inter

    return inter / union if union > 0 else 0.0


def deduplicate_by_iou(detections, iou_threshold=0.9):
    """Keep one detection per overlapping group (IoU >= iou_threshold). Prefer higher confidence."""
    if not detections:
        return []

    # Sort by confidence descending; keep first, then add only if no kept box overlaps >= threshold
    sorted_d = sorted(detections, key=lambda x: -x["conf"])
    kept = []

    for d in sorted_d:
        if not any(box_iou(d["box"], k["box"]) >= iou_threshold for k in kept):
            kept.append(d)

    return kept


def parse_args():
    p = argparse.ArgumentParser(
        description="D-FINE (person/car) → group → Jina-CLIP-v2 on crops inside groups"
    )
    p.add_argument("--refs", required=True, help="Reference images folder for Jina (e.g. refs/)")
    p.add_argument("--input", required=True, help="Full-frame images folder")
    p.add_argument("--output", default="pipeline_results", help="Output folder (CSV, etc.)")
    p.add_argument("--det-threshold", type=float, default=0.13, help="D-FINE score threshold")
    p.add_argument("--group-dist", type=float, default=None, help="Group distance (default: 0.1 * max(H,W))")
    p.add_argument("--min-side", type=int, default=40, help="Min side of expanded bbox in px (skip smaller)")
    p.add_argument("--crop-dedup-iou", type=float, default=0.35, help="Min IoU to treat two crops as same object (keep larger)")
    p.add_argument("--no-squarify", action="store_true", help="Skip squarify; use expanded bbox only (tighter crops, often better recognition)")
    p.add_argument("--padding", type=float, default=0.2, help="Crop padding around group box (0.2 = 20%%)")
    p.add_argument("--conf-threshold", type=float, default=0.75, help="Jina accept confidence")
    p.add_argument("--gap-threshold", type=float, default=0.05, help="Jina accept gap")
    p.add_argument("--text-weight", type=float, default=0.3)
    p.add_argument("--max-images", type=int, default=None)
    p.add_argument("--device", default=None)
    p.add_argument("--dfine-model", choices=list(DFINE_MODEL_IDS.keys()), default="large-obj365", help="D-FINE model")
    return p.parse_args()


def get_person_car_label_ids(model):
    """Return set of label IDs for person and car (Objects365: Person, Car, SUV, etc.)."""
    id2label = getattr(model.config, "id2label", None) or {}
    ids = set()

    for idx, name in id2label.items():
        try:
            i = int(idx)
        except (ValueError, TypeError):
            continue

        n = (name or "").lower()
        if "person" in n or n in ("car", "suv"):
            ids.add(i)

    return ids


def run_dfine(image, processor, model, device, score_threshold):
    """Run D-FINE, return all detections as list of {box, score, label_id, label}."""
    from PIL import Image

    if isinstance(image, Image.Image):
        pil = image.convert("RGB")
    else:
        pil = Image.fromarray(image).convert("RGB")

    w, h = pil.size
    target_size = torch.tensor([[h, w]], device=device)

    inputs = processor(images=pil, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}

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

    target_sizes = target_size.to(outputs["logits"].device)
    results = processor.post_process_object_detection(
        outputs,
        target_sizes=target_sizes,
        threshold=score_threshold,
    )

    id2label = getattr(model.config, "id2label", {}) or {}
    detections = []

    for result in results:
        for score, label_id, box in zip(
            result["scores"],
            result["labels"],
            result["boxes"]
        ):
            sid = int(label_id.item())
            detections.append({
                "box": [float(x) for x in box.cpu().tolist()],
                "conf": float(score.item()),
                "cls": sid,
                "label": id2label.get(sid, str(sid)),
            })

    return detections


def main():
    args = parse_args()
    device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")

    input_dir = Path(args.input)
    output_dir = Path(args.output)
    refs_dir = Path(args.refs)
    output_dir.mkdir(parents=True, exist_ok=True)

    if not refs_dir.is_dir():
        raise SystemExit(f"Refs folder not found: {refs_dir}")
    if not input_dir.is_dir():
        raise SystemExit(f"Input folder not found: {input_dir}")

    paths = sorted(
        p for p in input_dir.iterdir()
        if p.suffix.lower() in IMAGE_EXTS
    )
    if args.max_images is not None:
        paths = paths[: args.max_images]

    if not paths:
        raise SystemExit(f"No images in {input_dir}")

    # Load D-FINE
    dfine_model_id = DFINE_MODEL_IDS.get(args.dfine_model, DFINE_MODEL_IDS["large-obj365"])
    print(f"[*] Loading D-FINE ({dfine_model_id})...")
    t0 = time.perf_counter()
    image_processor = AutoImageProcessor.from_pretrained(dfine_model_id)
    dfine_model = DFineForObjectDetection.from_pretrained(dfine_model_id)
    dfine_model = dfine_model.to(device).eval()
    person_car_ids = get_person_car_label_ids(dfine_model)
    print(f"  Person/car label IDs: {person_car_ids} ({time.perf_counter()-t0:.1f}s)")

    # Load Jina-CLIP-v2 + build refs
    print("[*] Loading Jina-CLIP-v2 and building refs...")
    t0 = time.perf_counter()
    jina_encoder = JinaCLIPv2Encoder(device)
    ref_labels, ref_embs = build_refs(
        jina_encoder,
        refs_dir,
        TRUNCATE_DIM,
        args.text_weight,
        batch_size=16
    )
    print(f"  Jina refs: {ref_labels} ({time.perf_counter()-t0:.1f}s)\n")

    jina_crops_dir = output_dir / "jina_crops"
    jina_crops_dir.mkdir(parents=True, exist_ok=True)

    # CSV
    csv_path = output_dir / "results.csv"
    f = open(csv_path, "w", newline="")
    w = csv.writer(f)
    w.writerow([
        "image",
        "crop_filename",
        "group_idx",
        "crop_x1",
        "crop_y1",
        "crop_x2",
        "crop_y2",
        "bbox_x1",
        "bbox_y1",
        "bbox_x2",
        "bbox_y2",
        "dfine_label",
        "dfine_conf",
        "jina_prediction",
        "jina_confidence",
        "jina_status",
    ])

    for img_path in paths:
        pil = Image.open(img_path).convert("RGB")
        img_w, img_h = pil.size
        group_dist = args.group_dist if args.group_dist is not None else 0.1 * max(img_h, img_w)

        # 1) D-FINE: detect everything, keep all bboxes for the image
        detections = run_dfine(
            pil,
            image_processor,
            dfine_model,
            device,
            args.det_threshold
        )

        person_car = [d for d in detections if d["cls"] in person_car_ids and d["conf"] > PERSON_CAR_MIN_CONF]
        if not person_car:
            continue

        # 2) Group person/car detections (same as reference)
        grouped = group_detections(person_car, group_dist)
        grouped.sort(key=lambda x: x["conf"], reverse=True)
        top_groups = grouped[:10]  # limit groups per image

        # 3) Collect all candidate crops (bboxes inside person/car groups)
        # Each: (crop_box, crop_pil, d, gidx, crop_idx, x1, y1, x2, y2)
        candidates = []

        for gidx, grp in enumerate(top_groups):
            x1, y1, x2, y2 = grp["box"]
            group_box = [x1, y1, x2, y2]
            group_box_with_margin = expand_box_by_margin(group_box, PERSON_CAR_GROUP_MARGIN, img_w, img_h)

            inside = [
                d for d in detections
                if box_center_inside(d["box"], group_box_with_margin) and d["cls"] not in person_car_ids
            ]
            inside = deduplicate_by_iou(inside, iou_threshold=0.9)

            for crop_idx, d in enumerate(inside):
                bx1, by1, bx2, by2 = [float(x) for x in d["box"]]
                obj_w, obj_h = bx2 - bx1, by2 - by1
                if obj_w <= 0 or obj_h <= 0:
                    continue

                # Small objects (min side < 24 px): expand by 60%; larger: 30%
                min_side_obj = min(obj_w, obj_h)
                pad_ratio = 0.6 if min_side_obj < 24 else 0.3
                pad_x = obj_w * pad_ratio
                pad_y = obj_h * pad_ratio
                bx1 = max(0, int(bx1 - pad_x))
                by1 = max(0, int(by1 - pad_y))
                bx2 = min(img_w, int(bx2 + pad_x))
                by2 = min(img_h, int(by2 + pad_y))

                if bx2 <= bx1 or by2 <= by1:
                    continue

                if min(bx2 - bx1, by2 - by1) < args.min_side:
                    continue

                expanded_box = [bx1, by1, bx2, by2]
                candidates.append((expanded_box, d, gidx, crop_idx, x1, y1, x2, y2))

        # 4) Dedup on EXPANDED boxes (before squarify), keep larger; then squarify only kept
        def crop_area(box):
            return (box[2] - box[0]) * (box[3] - box[1])

        candidates.sort(key=lambda c: -crop_area(c[0]))
        kept = []

        for c in candidates:
            expanded_box = c[0]

            def is_same_object(box_a, box_b):
                if box_iou(box_a, box_b) >= args.crop_dedup_iou:
                    return True
                if box_center_inside(box_a, box_b) or box_center_inside(box_b, box_a):
                    return True
                return False

            if not any(is_same_object(expanded_box, k[0]) for k in kept):
                kept.append(c)

        # 5) Optionally squarify, then run Jina on kept crops
        for i, (expanded_box, d, gidx, crop_idx, x1, y1, x2, y2) in enumerate(kept):
            if not args.no_squarify:
                bx1, by1, bx2, by2 = squarify_crop_box(
                    expanded_box[0],
                    expanded_box[1],
                    expanded_box[2],
                    expanded_box[3],
                    img_w,
                    img_h
                )
            else:
                bx1, by1, bx2, by2 = expanded_box[0], expanded_box[1], expanded_box[2], expanded_box[3]

            crop_pil = pil.crop((bx1, by1, bx2, by2))
            crop_name = f"{img_path.stem}_g{gidx}_{i}_{bx1}_{by1}_{bx2}_{by2}{img_path.suffix}"

            q_jina = jina_encoder.encode_images([crop_pil], TRUNCATE_DIM)
            result_jina = jina_classify(
                q_jina,
                ref_labels,
                ref_embs,
                args.conf_threshold,
                args.gap_threshold
            )

            if result_jina["prediction"] in ref_labels:
                label_jina = result_jina["prediction"]
                conf_jina = result_jina["confidence"]
            else:
                label_jina = f"unnamed (dfine: {d['label']})"
                conf_jina = 0.0

            ann_jina = draw_label_on_image(crop_pil, label_jina, conf_jina)
            ann_jina.save(jina_crops_dir / crop_name)

            w.writerow([
                img_path.name,
                crop_name,
                gidx,
                x1,
                y1,
                x2,
                y2,
                bx1,
                by1,
                bx2,
                by2,
                d["label"],
                f"{d['conf']:.4f}",
                result_jina["prediction"],
                f"{result_jina['confidence']:.4f}",
                result_jina["status"],
            ])

    f.close()
    print(f"[*] Wrote {csv_path}")
    print(f"[*] Jina crops: {jina_crops_dir}")


# -----------------------------------------------------------------------------
# Single-image runner for Gradio app: D-FINE first, then Jina
# -----------------------------------------------------------------------------

_APP_DFINE = None  # (model_id, image_processor, dfine_model, person_car_ids)
_APP_CLASSIFIERS = {}  # {classifier_name: (classifier_instance, refs_dir_str)}

DFINE_MODEL_IDS = {
    # obj365
    "small-obj365": "ustc-community/dfine-small-obj365",
    "medium-obj365": "ustc-community/dfine-medium-obj365",
    "large-obj365": "ustc-community/dfine-large-obj365",
    # coco
    "small-coco": "ustc-community/dfine-small-coco",
    "medium-coco": "ustc-community/dfine-medium-coco",
    "large-coco": "ustc-community/dfine-large-coco",
    # obj2coco
    "small-obj2coco": "ustc-community/dfine-small-obj2coco",
    "medium-obj2coco": "ustc-community/dfine-medium-obj2coco",
    "large-obj2coco": "ustc-community/dfine-large-obj2coco-e25",
}

CLASSIFIER_CHOICES = ["jina", "siglip-224", "siglip-256", "siglip-384", "siglip2_onnx"]


def _load_classifier(classifier_name, device, refs_dir=None, labels=None):
    """Factory: load and initialize a classifier by name."""
    if refs_dir:
        refs_dir = Path(refs_dir)

    if classifier_name == "jina":
        jina_encoder = JinaCLIPv2Encoder(device)
        ref_labels, ref_embs = build_refs(jina_encoder, refs_dir, TRUNCATE_DIM, 0.3, batch_size=16)
        return ("jina_wrapped", jina_encoder, ref_labels, ref_embs)

    if classifier_name.startswith("siglip-"):
        from siglip_zeroshot import SigLIPClassifier, SIGLIP_MODELS
        if classifier_name not in SIGLIP_MODELS:
            raise ValueError(f"Unknown SigLIP model: {classifier_name}. Choose from {list(SIGLIP_MODELS.keys())}")
        clf = SigLIPClassifier(device, model_key=classifier_name)
        clf.build_refs(refs_dir=refs_dir, labels=labels)
        return clf

    if classifier_name == "siglip2_onnx":
        from siglip2_onnx_zeroshot import SigLIP2ONNXClassifier
        clf = SigLIP2ONNXClassifier(device)
        clf.build_refs(refs_dir=refs_dir, labels=labels)
        return clf

    raise ValueError(f"Unknown classifier: {classifier_name}. Choose from {CLASSIFIER_CHOICES}")


def _classify_crop(classifier, crop, conf_threshold, gap_threshold):
    """Unified classify call that works for both Jina (tuple) and SigLIP-style classifiers."""
    if isinstance(classifier, tuple) and classifier[0] == "jina_wrapped":
        _, jina_encoder, ref_labels, ref_embs = classifier
        q = jina_encoder.encode_images([crop], TRUNCATE_DIM)
        return jina_classify(q, ref_labels, ref_embs, conf_threshold, gap_threshold)
    else:
        return classifier.classify_crop(crop, conf_threshold, gap_threshold)


def run_single_image(
    pil_image,
    refs_dir=None,
    device=None,
    dfine_model="large",
    det_threshold=0.3,
    conf_threshold=0.75,
    gap_threshold=0.05,
    min_side=40,
    crop_dedup_iou=0.35,
    squarify=True,
    min_display_conf=None,
    classifier="siglip-256",
    labels=None,
):
    """
    Run D-FINE on one image, then classify small-object crops.

    refs_dir: path to refs folder (str or Path), optional if labels provided.
    labels: list of class label strings for zero-shot classifiers.
    dfine_model: key from DFINE_MODEL_IDS.

    Returns (group_crop_images, known_crop_composites, status_message).
    """
    import numpy as np

    if min_display_conf is None:
        min_display_conf = MIN_DISPLAY_CONF
    from PIL import Image

    global _APP_DFINE

    if refs_dir:
        refs_dir = Path(refs_dir)
        if not refs_dir.is_dir():
            return [], [], f"Refs folder not found: {refs_dir}"

    if not refs_dir and not labels:
        return [], [], "Provide either refs_dir or labels."

    dfine_model = (dfine_model or "large-obj365").strip().lower()
    if dfine_model not in DFINE_MODEL_IDS:
        dfine_model = "large-obj365"
    model_id = DFINE_MODEL_IDS[dfine_model]

    device = device or ("cuda" if torch.cuda.is_available() else "cpu")
    print(f"[*] Device: {device}")

    pil = pil_image.convert("RGB") if isinstance(pil_image, Image.Image) else Image.fromarray(pil_image).convert("RGB")
    img_w, img_h = pil.size
    group_dist = 0.1 * max(img_h, img_w)

    # Load D-FINE (reload if user switched model)
    if _APP_DFINE is None or _APP_DFINE[0] != dfine_model:
        print(f"[*] Loading D-FINE ({model_id})...")
        image_processor = AutoImageProcessor.from_pretrained(model_id)
        dfine_model_obj = DFineForObjectDetection.from_pretrained(model_id)
        dfine_model_obj = dfine_model_obj.to(device).eval()
        person_car_ids = get_person_car_label_ids(dfine_model_obj)
        _APP_DFINE = (dfine_model, image_processor, dfine_model_obj, person_car_ids)

    _model_id, image_processor, dfine_model_obj, person_car_ids = _APP_DFINE

    # Apply user's D-FINE detection threshold to the chosen model (medium or large)
    detections = run_dfine(pil, image_processor, dfine_model_obj, device, det_threshold)
    person_car = [d for d in detections if d["cls"] in person_car_ids and d["conf"] > PERSON_CAR_MIN_CONF]
    if not person_car:
        return [], [], "No person/car detected (or none with confidence > 0.9). No small-object crops."

    grouped = group_detections(person_car, group_dist)
    grouped.sort(key=lambda x: x["conf"], reverse=True)
    top_groups = grouped[:10]

    # Load classifier
    global _APP_CLASSIFIERS
    cache_key = str(labels) if labels else str(refs_dir)
    clf_key = classifier
    if clf_key not in _APP_CLASSIFIERS or _APP_CLASSIFIERS[clf_key][1] != cache_key:
        clf_instance = _load_classifier(classifier, device, refs_dir=refs_dir, labels=labels)
        _APP_CLASSIFIERS[clf_key] = (clf_instance, cache_key)

    clf_instance = _APP_CLASSIFIERS[clf_key][0]

    results_per_crop = []
    group_crop_images = []
    classification_log = []

    # Non-person/car detections from the full-frame pass, reused per group
    other_detections = [d for d in detections if d["cls"] not in person_car_ids]

    # For each person/car group: crop (with 10% margin), reuse full-frame detections that fall inside, then classify
    for gidx, grp in enumerate(top_groups):
        group_box = [grp["box"][0], grp["box"][1], grp["box"][2], grp["box"][3]]
        crop_box = expand_box_by_margin(group_box, PERSON_CAR_GROUP_MARGIN, img_w, img_h)
        gx1 = max(0, int(crop_box[0]))
        gy1 = max(0, int(crop_box[1]))
        gx2 = min(img_w, int(crop_box[2]))
        gy2 = min(img_h, int(crop_box[3]))
        if gx2 <= gx1 or gy2 <= gy1:
            continue
        crop_pil = pil.crop((gx1, gy1, gx2, gy2)).copy().convert("RGB")
        crop_w, crop_h = crop_pil.size

        # Filter full-frame detections whose center falls inside this crop box, then remap to crop-local coords
        inside_full = [d for d in other_detections if box_center_inside(d["box"], [gx1, gy1, gx2, gy2])]
        inside = []
        for d in inside_full:
            remapped = dict(d)
            fx1, fy1, fx2, fy2 = d["box"]
            remapped["box"] = [
                max(0, fx1 - gx1),
                max(0, fy1 - gy1),
                min(crop_w, fx2 - gx1),
                min(crop_h, fy2 - gy1),
            ]
            inside.append(remapped)
        inside = deduplicate_by_iou(inside, iou_threshold=0.9)

        candidates = []
        for d in inside:
            bx1, by1, bx2, by2 = [float(x) for x in d["box"]]
            obj_w, obj_h = bx2 - bx1, by2 - by1
            if obj_w <= 0 or obj_h <= 0:
                continue
            min_side_obj = min(obj_w, obj_h)
            pad_ratio = 0.6 if min_side_obj < 24 else 0.3
            pad_x = obj_w * pad_ratio
            pad_y = obj_h * pad_ratio
            bx1 = max(0.0, bx1 - pad_x)
            by1 = max(0.0, by1 - pad_y)
            bx2 = min(crop_w, bx2 + pad_x)
            by2 = min(crop_h, by2 + pad_y)
            if bx2 <= bx1 or by2 <= by1:
                continue
            w, h = bx2 - bx1, by2 - by1
            if min(w, h) < MIN_OBJECT_CROP_SIDE:
                need = MIN_OBJECT_CROP_SIDE - min(w, h)
                half = need / 2.0
                if w < h:
                    bx1 = max(0, bx1 - half)
                    bx2 = min(crop_w, bx2 + half)
                else:
                    by1 = max(0, by1 - half)
                    by2 = min(crop_h, by2 + half)
                w, h = bx2 - bx1, by2 - by1
                if w < MIN_OBJECT_CROP_SIDE:
                    add = (MIN_OBJECT_CROP_SIDE - w) / 2
                    bx1 = max(0, bx1 - add)
                    bx2 = min(crop_w, bx2 + add)
                if h < MIN_OBJECT_CROP_SIDE:
                    add = (MIN_OBJECT_CROP_SIDE - h) / 2
                    by1 = max(0, by1 - add)
                    by2 = min(crop_h, by2 + add)
            bx1, by1, bx2, by2 = int(bx1), int(by1), int(bx2), int(by2)
            if bx2 <= bx1 or by2 <= by1:
                continue
            candidates.append(([bx1, by1, bx2, by2], d, gidx))

        def crop_area(box):
            return (box[2] - box[0]) * (box[3] - box[1])

        candidates.sort(key=lambda c: -crop_area(c[0]))
        kept = []
        for c in candidates:
            expanded_box = c[0]
            if not any(
                box_iou(expanded_box, k[0]) >= crop_dedup_iou
                or box_center_inside(expanded_box, k[0])
                or box_center_inside(k[0], expanded_box)
                for k in kept
            ):
                kept.append(c)

        for (bx1, by1, bx2, by2), d, _ in kept:
            if squarify:
                bx1, by1, bx2, by2 = squarify_crop_box(bx1, by1, bx2, by2, crop_w, crop_h)
            small_crop = crop_pil.crop((bx1, by1, bx2, by2))
            result = _classify_crop(clf_instance, small_crop, conf_threshold, gap_threshold)
            raw_pred = result["prediction"]
            pred = raw_pred if raw_pred != "unknown" else f"unknown ({d['label']})"
            conf = result["confidence"]
            results_per_crop.append((gidx, (bx1, by1, bx2, by2), small_crop, pred, conf))

            # Build per-crop log line
            sims_str = ", ".join(f"{k}: {v:.4f}" for k, v in result.get("all_sims", {}).items())
            classification_log.append(
                f"[group {gidx}] dfine: {d['label']} ({d['conf']:.3f}) → "
                f"{pred} (conf={conf:.4f}, gap={result['gap']:.4f}, 2nd={result.get('second_best','?')}) "
                f"| {result['status']} | {sims_str}"
            )

        # Draw bboxes on this group crop (bboxes already in crop coords)
        boxes_to_draw = [
            (bx1, by1, bx2, by2, pred, conf)
            for (gidx2, (bx1, by1, bx2, by2), _sc, pred, conf) in results_per_crop
            if gidx2 == gidx
        ]
        if boxes_to_draw:
            crop_pil_drawn = draw_bboxes_on_image(crop_pil.copy(), boxes_to_draw)
        else:
            crop_pil_drawn = crop_pil
        group_crop_images.append(np.array(crop_pil_drawn))

    log_text = f"Classifier: {classifier} | {len(results_per_crop)} crops classified\n"
    log_text += "\n".join(classification_log) if classification_log else "(no crops)"

    if not results_per_crop:
        return group_crop_images if group_crop_images else [], [], log_text + "\nNo small-object crops: no object detections (gun/phone/etc.) found inside person/car groups, or all were below min size."

    # Build known-only gallery: only objects with conf >= min_display_conf
    known_crop_composites = []
    for (_gidx, _box, crop_pil, pred, conf) in results_per_crop:
        if pred.startswith("unknown") or conf < min_display_conf:
            continue
        composite = draw_label_on_image(crop_pil, pred, conf)
        known_crop_composites.append(np.array(composite))

    return group_crop_images, known_crop_composites, log_text


if __name__ == "__main__":
    main()