Detectron2: Custom Data Augmentation Implementation

Question

I am working on an underwater image detection problem using detection2. I have applied an image enhancement augmentation offline (by storing the newly processed data in a separate folder). But I need to apply the image enhancement on the fly, such that it is fed into the model during training time (with some probability). I tried integrating this module on the fly but it's taking too much time. Can anybody help me to resolve this issue? I am using RetinaNet as a backbone. I tried the following:

def map_enhance(dataset_dict):
    dataset_dict = copy.deepcopy(dataset_dict)  # it will be modified by code below
    # can use other ways to read image
    image = utils.read_image(dataset_dict["file_name"], format="BGR")
    # Apply some image enhancement technique:

    # Create the format that the model expects:
    return {"image": image,
        "instances": utils.annotations_to_instances(annos, image.shape[1:]),
        "width": image.shape[1],
        "height": image.shape[2]}


class TrainerEnhance(DefaultTrainer):
    @classmethod
    def build_train_loader(cls, cfg):
        if cfg.INPUT.ENHANCE_IMAGE:
            mapper = DatasetMapper(cfg, is_train=True, augmentations=[map_enhance])
        else:
            mapper = None
        return build_detection_train_loader(cfg, mapper=mapper)

But unfortunately, that did not work!

Detectron2: Custom Data Augmentation Implementation

0 Answers0