Spaces:
Sleeping
Sleeping
File size: 17,091 Bytes
82551bb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 | """
Find best conf_threshold and gap_threshold for Jina and Nomic using COCO ground truth.
Expects full_frames/ with images and annotations.coco.json (COCO format).
Runs the same detection + crop pipeline, matches each crop to a GT annotation (IoU),
then grid-searches (conf_threshold, gap_threshold) to maximize accuracy.
"""
import argparse
import json
from pathlib import Path
import numpy as np
import torch
from PIL import Image
from transformers import AutoImageProcessor, DFineForObjectDetection
from dfine_jina_pipeline import (
box_center_inside,
box_iou,
deduplicate_by_iou,
get_person_car_label_ids,
group_detections,
run_dfine,
squarify_crop_box,
)
from jina_fewshot import IMAGE_EXTS, TRUNCATE_DIM, JinaCLIPv2Encoder, build_refs, draw_label_on_image
from nomic_fewshot import NomicTextEncoder, NomicVisionEncoder, build_refs_nomic
# Our 4 classes (same order as refs)
CLASS_NAMES = ["cigarette", "gun", "knife", "phone"]
def coco_bbox_to_xyxy(bbox):
"""COCO bbox [x, y, w, h] -> [x1, y1, x2, y2]. Tolerate string numbers from JSON."""
x, y, w, h = (float(v) for v in bbox)
return [x, y, x + w, y + h]
def map_category_to_class(name: str) -> str | None:
"""Map COCO category name to one of our 4 classes, or None if other."""
n = (name or "").strip().lower()
if "cigarette" in n:
return "cigarette"
if any(x in n for x in ("gun", "pistol", "handgun", "firearm")):
return "gun"
if "knife" in n or "blade" in n:
return "knife"
if any(x in n for x in ("phone", "cell", "mobile", "smartphone", "telephone")):
return "phone"
return None
def load_coco_gt(annotations_path: Path):
"""
Load COCO JSON. Returns:
- file_to_gts: dict[file_name] = list of (bbox_xyxy, category_name)
- categories: list of category dicts from COCO
"""
with open(annotations_path) as f:
data = json.load(f)
images = {im["id"]: im for im in data.get("images", [])}
categories = {c["id"]: c["name"] for c in data.get("categories", [])}
file_to_gts = {}
for im in images.values():
file_to_gts[im["file_name"]] = []
for ann in data.get("annotations", []):
image_id = ann["image_id"]
cat_name = categories.get(ann["category_id"], "")
bbox_xyxy = coco_bbox_to_xyxy(ann["bbox"])
file_name = images[image_id]["file_name"]
file_to_gts[file_name].append((bbox_xyxy, cat_name))
# Also index by basename for lookup by Path.name
by_basename = {}
for fn, gts in file_to_gts.items():
by_basename[Path(fn).name] = gts
return by_basename, data.get("categories", [])
def assign_gt_to_crop(crop_box_xyxy, gt_list, iou_min=0.3):
"""
Find best overlapping GT for this crop. Returns (gt_class or None, iou).
gt_class is one of CLASS_NAMES (mapped from category).
"""
best_iou = 0.0
best_class = None
for bbox_xyxy, cat_name in gt_list:
iou = box_iou(crop_box_xyxy, bbox_xyxy)
if iou >= iou_min and iou > best_iou:
cls = map_category_to_class(cat_name)
if cls is not None:
best_iou = iou
best_class = cls
return best_class, best_iou
def parse_args():
p = argparse.ArgumentParser(description="Tune Jina/Nomic thresholds using COCO GT")
p.add_argument("--input", default="full_frames", help="Folder with images and annotations.coco.json")
p.add_argument("--annotations", default=None, help="Path to annotations.coco.json (default: input/_annotations.coco.json)")
p.add_argument("--refs", required=True, help="Reference images folder (for Jina + Nomic refs)")
p.add_argument("--output", default="threshold_tuning", help="Output folder for results CSV")
p.add_argument("--det-threshold", type=float, default=0.3)
p.add_argument("--group-dist", type=float, default=None)
p.add_argument("--expand", type=float, default=0.3)
p.add_argument("--min-side", type=int, default=40)
p.add_argument("--text-weight", type=float, default=0.3)
p.add_argument("--iou-min", type=float, default=0.3, help="Min IoU to match crop to GT")
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("--max-images", type=int, default=None)
p.add_argument("--device", default=None)
p.add_argument("--no-save-crops", action="store_true", help="Do not save annotated crop images")
p.add_argument("--save-conf", type=float, default=0.5, help="Conf threshold for saved crop labels")
p.add_argument("--save-gap", type=float, default=0.02, help="Gap threshold for saved crop labels")
return p.parse_args()
def main():
args = parse_args()
device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")
input_dir = Path(args.input)
refs_dir = Path(args.refs)
output_dir = Path(args.output)
output_dir.mkdir(parents=True, exist_ok=True)
annotations_path = Path(args.annotations) if args.annotations else input_dir / "_annotations.coco.json"
if not annotations_path.is_file():
raise SystemExit(f"Annotations not found: {annotations_path}")
file_to_gts, _ = load_coco_gt(annotations_path)
print(f"[*] Loaded GT for {len(file_to_gts)} images from {annotations_path}")
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]
# Only images that appear in COCO
paths = [p for p in paths if p.name in file_to_gts]
if not paths:
raise SystemExit("No images in input that have COCO annotations.")
print(f"[*] Loading D-FINE...")
image_processor = AutoImageProcessor.from_pretrained("ustc-community/dfine-medium-obj365")
dfine_model = DFineForObjectDetection.from_pretrained("ustc-community/dfine-medium-obj365")
dfine_model = dfine_model.to(device).eval()
person_car_ids = get_person_car_label_ids(dfine_model)
print("[*] Loading Jina-CLIP-v2 and building refs...")
jina_encoder = JinaCLIPv2Encoder(device)
ref_labels, ref_embs = build_refs(
jina_encoder, refs_dir, TRUNCATE_DIM, args.text_weight, batch_size=16
)
assert ref_labels == CLASS_NAMES, f"Ref order {ref_labels}"
print("[*] Loading Nomic (vision + text) and building refs (same as Jina: text_weight 0.3)...")
nomic_encoder = NomicVisionEncoder(device)
nomic_text_encoder = NomicTextEncoder(device)
ref_labels_nomic, ref_embs_nomic = build_refs_nomic(
nomic_encoder, refs_dir, batch_size=16,
text_encoder=nomic_text_encoder, text_weight=args.text_weight,
)
# Optional: save annotated crops for Jina and Nomic, raw crops (no label), and person/car grouping crops
save_crops = not args.no_save_crops
if save_crops:
jina_crops_dir = output_dir / "jina_crops"
nomic_crops_dir = output_dir / "nomic_crops"
crops_no_label_dir = output_dir / "crops"
detection_crops_dir = output_dir / "detection_crops"
jina_crops_dir.mkdir(parents=True, exist_ok=True)
nomic_crops_dir.mkdir(parents=True, exist_ok=True)
crops_no_label_dir.mkdir(parents=True, exist_ok=True)
detection_crops_dir.mkdir(parents=True, exist_ok=True)
# Collect per-crop: gt_class, jina sims/conf/gap, nomic sims/conf/gap (only crops with gt in our 4 classes)
rows = []
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)
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]
if not person_car:
continue
grouped = group_detections(person_car, group_dist)
grouped.sort(key=lambda x: x["conf"], reverse=True)
gt_list = file_to_gts.get(img_path.name, [])
if not gt_list:
continue
# 1) Collect all candidate crops (bboxes inside person/car groups, with GT match)
# Each candidate: (crop_box, crop_pil, gt_class, gidx, crop_idx)
candidates = []
for gidx, grp in enumerate(grouped[:10]):
x1, y1, x2, y2 = grp["box"]
group_box = [x1, y1, x2, y2]
# Save person/car grouping output crop (detection crop only)
if save_crops:
gx1 = max(0, int(x1))
gy1 = max(0, int(y1))
gx2 = min(img_w, int(x2))
gy2 = min(img_h, int(y2))
if gx2 > gx1 and gy2 > gy1:
group_crop = pil.crop((gx1, gy1, gx2, gy2))
group_crop.save(detection_crops_dir / f"{img_path.stem}_group{gidx}.jpg")
inside = [
d for d in detections
if box_center_inside(d["box"], group_box)
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
pad_x, pad_y = obj_w * args.expand, obj_h * args.expand
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]
gt_class, _ = assign_gt_to_crop(expanded_box, gt_list, args.iou_min)
if gt_class is None:
continue
candidates.append((expanded_box, gt_class, gidx, crop_idx))
# 2) 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])) # largest first
kept = []
for c in candidates:
expanded_box = c[0]
# Skip if same object: IoU above threshold, or one box's center is inside the other
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)
# 3) Optionally squarify, then run Jina/Nomic only on kept crops
for i, (expanded_box, gt_class, gidx, crop_idx) 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_box = [bx1, by1, bx2, by2]
crop_pil = pil.crop((bx1, by1, bx2, by2))
bbox_suffix = f"_{bx1}_{by1}_{bx2}_{by2}"
crop_name = f"{img_path.stem}_g{gidx}_{i}{bbox_suffix}.jpg"
q_jina = jina_encoder.encode_images([crop_pil], TRUNCATE_DIM)
sims_jina = (q_jina @ ref_embs.T).squeeze(0)
best_jina = int(np.argmax(sims_jina))
conf_jina = float(sims_jina[best_jina])
gap_jina = float(sims_jina[best_jina] - np.partition(sims_jina, -2)[-2])
q_nomic = nomic_encoder.encode_images([crop_pil])
sims_nomic = (q_nomic @ ref_embs_nomic.T).squeeze(0)
best_nomic = int(np.argmax(sims_nomic))
conf_nomic = float(sims_nomic[best_nomic])
gap_nomic = float(sims_nomic[best_nomic] - np.partition(sims_nomic, -2)[-2])
rows.append({
"gt": gt_class,
"jina_best_idx": best_jina,
"jina_conf": conf_jina,
"jina_gap": gap_jina,
"nomic_best_idx": best_nomic,
"nomic_conf": conf_nomic,
"nomic_gap": gap_nomic,
})
if save_crops:
crop_pil.save(crops_no_label_dir / crop_name)
sc, sg = args.save_conf, args.save_gap
label_jina = ref_labels[best_jina] if (conf_jina >= sc and gap_jina >= sg) else f"unknown (gt:{gt_class})"
label_nomic = ref_labels_nomic[best_nomic] if (conf_nomic >= sc and gap_nomic >= sg) else f"unknown (gt:{gt_class})"
ann_jina = draw_label_on_image(crop_pil, label_jina, conf_jina)
ann_nomic = draw_label_on_image(crop_pil, label_nomic, conf_nomic)
ann_jina.save(jina_crops_dir / crop_name)
ann_nomic.save(nomic_crops_dir / crop_name)
if not rows:
raise SystemExit("No crops matched to GT (with our 4 classes). Check annotations and iou_min.")
print(f"[*] {len(rows)} crops with GT in {{cigarette, gun, knife, phone}}")
if save_crops:
print(f"[*] Annotated crops saved to {jina_crops_dir} and {nomic_crops_dir}")
print(f"[*] Raw crops (no label) saved to {crops_no_label_dir}")
print(f"[*] Person/car grouping crops saved to {detection_crops_dir}")
# Grid search
conf_candidates = [0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80]
gap_candidates = [0.02, 0.05, 0.08, 0.10]
def accuracy_jina(conf_t, gap_t):
correct = 0
for r in rows:
pred = ref_labels[r["jina_best_idx"]] if (r["jina_conf"] >= conf_t and r["jina_gap"] >= gap_t) else "unknown"
if pred == r["gt"]:
correct += 1
return correct / len(rows)
def accuracy_nomic(conf_t, gap_t):
correct = 0
for r in rows:
pred = ref_labels_nomic[r["nomic_best_idx"]] if (r["nomic_conf"] >= conf_t and r["nomic_gap"] >= gap_t) else "unknown"
if pred == r["gt"]:
correct += 1
return correct / len(rows)
best_jina_acc = -1
best_jina_conf = best_jina_gap = None
for c in conf_candidates:
for g in gap_candidates:
acc = accuracy_jina(c, g)
if acc > best_jina_acc:
best_jina_acc = acc
best_jina_conf, best_jina_gap = c, g
best_nomic_acc = -1
best_nomic_conf = best_nomic_gap = None
for c in conf_candidates:
for g in gap_candidates:
acc = accuracy_nomic(c, g)
if acc > best_nomic_acc:
best_nomic_acc = acc
best_nomic_conf, best_nomic_gap = c, g
# Report
report_path = output_dir / "best_thresholds.txt"
with open(report_path, "w") as f:
f.write(f"Based on {len(rows)} crops with GT in {{cigarette, gun, knife, phone}}\n")
if save_crops:
f.write(f"Annotated crops: jina_crops/ and nomic_crops/ (conf>={args.save_conf}, gap>={args.save_gap})\n")
f.write("Raw crops (no label): crops/\n")
f.write("Person/car grouping only: detection_crops/\n")
f.write("\n")
f.write("Jina (best accuracy):\n")
f.write(f" conf_threshold = {best_jina_conf}\n")
f.write(f" gap_threshold = {best_jina_gap}\n")
f.write(f" accuracy = {best_jina_acc:.4f}\n\n")
f.write("Nomic (best accuracy):\n")
f.write(f" conf_threshold = {best_nomic_conf}\n")
f.write(f" gap_threshold = {best_nomic_gap}\n")
f.write(f" accuracy = {best_nomic_acc:.4f}\n")
print(f"\n[*] Best thresholds written to {report_path}")
print("\nJina best: conf_threshold={}, gap_threshold={} -> accuracy={:.4f}".format(
best_jina_conf, best_jina_gap, best_jina_acc))
print("Nomic best: conf_threshold={}, gap_threshold={} -> accuracy={:.4f}".format(
best_nomic_conf, best_nomic_gap, best_nomic_acc))
# Full grid CSV
import csv
csv_path = output_dir / "grid_search.csv"
with open(csv_path, "w", newline="") as f:
w = csv.writer(f)
w.writerow(["model", "conf_threshold", "gap_threshold", "accuracy"])
for c in conf_candidates:
for g in gap_candidates:
w.writerow(["jina", c, g, f"{accuracy_jina(c, g):.4f}"])
for c in conf_candidates:
for g in gap_candidates:
w.writerow(["nomic", c, g, f"{accuracy_nomic(c, g):.4f}"])
print(f"[*] Full grid written to {csv_path}")
if __name__ == "__main__":
main()
|