DShomin/arirang_detect

dacon arirang objecte detection

arirang_detect

dacon arirang objecte detection

공통수행 작업

1. download data from dacon
2. notebook/read_json.ipynb 를 실행하여 라벨을 csv로 바꿔준다.

YoloV5 사용법

1. yolo_v5 폴더 내에 python munge_data.py 으로 데이터 전처리 수행
2. python train.py을 이용하여 train
3. 예제 코드 python train.py --img 1024 --batch 1 --epochs 40 --data ./config/fold_0.yaml --cfg ./config/yolov5x.yaml --name yolov5x_fold0 --logdir fold0

EfficientDet 사용법

1. pretrained model download
2. python src/main.py

albumentation image augmentation 문제 발생

bbox가 0 ~ 1.0 구간을 벗어나면 error가 발생하는 문제
albumentations/augmentations/bbox_utils.py의 check_bbox를 아래와 같이 수정

def check_bbox(bbox):
    """Check if bbox boundaries are in range 0, 1 and minimums are lesser then maximums"""
    for name, value in zip(["x_min", "y_min", "x_max", "y_max"], bbox[:4]):
        if not 0 <= value <= 1:
            bbox = np.clip(bbox, 0., 1.)
            # raise ValueError(
            #     "Expected {name} for bbox {bbox} "
            #     "to be in the range [0.0, 1.0], got {value}.".format(bbox=bbox, name=name, value=value)
            # )
    x_min, y_min, x_max, y_max = bbox[:4]
    if x_max <= x_min:
        raise ValueError("x_max is less than or equal to x_min for bbox {bbox}.".format(bbox=bbox))
    if y_max <= y_min:
        raise ValueError("y_max is less than or equal to y_min for bbox {bbox}.".format(bbox=bbox))

Models

Variant	Download	mAP (val2017)	mAP (test-dev2017)	mAP (TF official val2017)	mAP (TF official test-dev2017)
lite0	tf_efficientdet_lite0.pth	32.0	TBD	N/A	N/A
D0	efficientdet_d0.pth	33.6	TBD	33.5	33.8
D0	tf_efficientdet_d0.pth	34.2	TBD	34.3	34.6
D1	efficientdet_d1.pth	39.4	39.5	39.1	39.6
D1	tf_efficientdet_d1.pth	40.1	TBD	40.2	40.5
D2	tf_efficientdet_d2.pth	43.4	TBD	42.5	43
D3	tf_efficientdet_d3.pth	47.1	TBD	47.2	47.5
D4	tf_efficientdet_d4.pth	49.2	TBD	49.3	49.7
D5	tf_efficientdet_d5.pth	51.2	TBD	51.2	51.5
D6	tf_efficientdet_d6.pth	52.0	TBD	52.1	52.6
D7	tf_efficientdet_d7.pth	53.1	53.4	53.4	53.7
D7X	tf_efficientdet_d7x.pth	54.3	TBD	54.4	55.1

NOTE: Official scores for all modules now using soft-nms, but still using normal NMS here.