plitip/team-qh-091-safesite-data-v2

Salamtak v2 detection dataset (team QH-091)

YOLO-format, 11 classes. train.zip -> images/train, labels/train; val.zip -> images/val, labels/val. Image names start with the source alias.

Classes

0 person, 1 helmet, 2 no_helmet, 3 vest, 4 no_vest, 5 machinery, 6 vehicle, 7 harness, 8 scaffold, 9 crane_hook, 10 utility_pole

Sources

alias	origin	license	ground-truth classes	pseudo-labelled classes	train	val
hh	keremberke/hard-hat-detection	CC BY 4.0	helmet, no_helmet	person, vest, no_vest, machinery, vehicle, utility_pole	11186	3962
cs	keremberke/construction-safety-object-detection	CC BY 4.0	person, helmet, no_helmet, vest, no_vest, machinery, vehicle	utility_pole	312	57
fr	Francesco/construction-safety-gsnvb	CC (Roboflow 100)	person, helmet, no_helmet, vest, no_vest	machinery, vehicle, utility_pole	897	90
chd	object-detection-qn97p/construction-hazard-detection	CC BY 4.0	person, helmet, no_helmet, vest, no_vest, machinery, vehicle, utility_pole	-	5820	669
ppe	Ultralytics Construction-PPE	CC BY 4.0	person, helmet, no_helmet, vest, no_vest, vehicle	machinery	800	200
hrn	project-xksuq/helmet-harness-detection	CC BY 4.0	helmet, no_helmet, harness	person, vest, no_vest, machinery, vehicle, utility_pole	566	152
scf	testscaff/hybrid-dataset-for-scaffold-detection	CC BY 4.0	scaffold	person, helmet, no_helmet, vest, no_vest, machinery, vehicle, utility_pole	3341	1250

Teacher (pseudo-labels)

• Model: yihong1120/Construction-Hazard-Detection, file models/yolo26/pt/yolo26x.pt (AGPL-3.0). Used only to label training images; it is not part of the product.
• Rule: on train images only, add teacher boxes with confidence >= 0.5 for classes that source does not label. A teacher helmet/no_helmet or vest/no_vest box overlapping (IoU >= 0.3) a ground-truth box of the opposite class is skipped; of two opposite teacher boxes only the more confident is kept. Ground-truth boxes are never deleted or moved. Validation labels are ground truth only.

De-duplication

{
  "phash_max_hamming": 4,
  "train_dropped_matching_val": 2753,
  "train_dropped_near_duplicates": 4000,
  "train_dropped_matching_val_by_source": {
    "fr": 207,
    "hh": 2438,
    "cs": 28,
    "hrn": 20,
    "chd": 59,
    "scf": 1
  },
  "train_dropped_near_duplicates_by_source": {
    "hh": 2159,
    "scf": 1658,
    "chd": 141,
    "hrn": 29,
    "fr": 12,
    "cs": 1
  }
}

Box counts

class	train GT	train pseudo	val GT
person	13392	41950	1724
helmet	36038	1789	10347
no_helmet	12152	58	2901
vest	6866	6688	963
no_vest	5077	23713	650
machinery	10945	2365	1214
vehicle	4093	549	468
harness	590	0	138
scaffold	3341	0	1250
crane_hook	0	0	0
utility_pole	1459	1074	186