farukcankaya

|          Method        |   AP   |  AP50  |  AP75  |  APs  |  APm   |  APl   |
|:----------------------:|:------:|:------:|:------:|:-----:|:------:|:------:|
|        Baseline        | 75.497 | 80.927 | 80.927 | 0.000 | 52.240 | 95.806 |
| InstanceColorJittering | 75.374 | 80.901 | 80.901 | 0.000 | 50.109 | 96.650 |
|       CopyPasteAug     | 78.176 | 83.978 | 83.978 | 0.000 | 56.604 | 96.981 |

farukcankaya

aug = InstanceColorJitterAugmentation(lambda img: ImageEnhance.Color(img).enhance(10))
augmentations = T.AugmentationList([aug])
aug_input = MultiModalAugInput(image, annos=d['annotations'])
transforms = augmentations(aug_input)
augmented_image = aug_input.image

farukcankaya

class InstanceColorJitterTransform(MultiModalTransform):
  def __init__(self, color_operation: Callable, instance_rate: float, min_count_to_apply: int) -> None:
    if not callable(color_operation):
        raise ValueError("color_operation parameter should be callable")
    super().__init__()
    self._set_attributes(locals())

  def apply_multi_modal(self, img, annos, *args):
    instance_count = len(annos)
    apply_count = max(self.min_count_to_apply, int(instance_count * self.instance_rate))

farukcankaya

def transform(self, tfm: Transform) -> None:
  if isinstance(tfm, MultiModalTransform):
      self.image, self.annos = \
          tfm.apply_multi_modal(self.image, self.annos)
  else:
      super().transform(tfm)

farukcankaya

class Transform:
  @abstractmethod
  def apply_image(self, img: np.ndarray):
    
  @abstractmethod
  def apply_coords(self, coords: np.ndarray):
    
  def apply_segmentation(self, segmentation: np.ndarray) -> np.ndarray:
    return self.apply_image(segmentation)
    

farukcankaya

class AugInput:
  def transform(self, tfm: Transform) -> None:
      self.image = tfm.apply_image(self.image)
      if self.boxes is not None:
          self.boxes = tfm.apply_box(self.boxes)
      if self.sem_seg is not None:
          self.sem_seg = tfm.apply_segmentation(self.sem_seg)

farukcankaya

class Augmentation:
  def __call__(self, aug_input) -> Transform:
    args = _get_aug_input_args(self, aug_input)
    tfm = self.get_transform(*args)
    aug_input.transform(tfm)
    return tfm

farukcankaya

# from: https://github.com/facebookresearch/detectron2/blob/998c4e15220a0db9f172b1e7ecf376a59b873f1d/detectron2/data/dataset_mapper.py#L163-L165
aug_input = T.AugInput(image, sem_seg=sem_seg_gt)
transforms = self.augmentations(aug_input)
image, sem_seg_gt = aug_input.image, aug_input.sem_seg

farukcankaya

farukcankaya

def do_train_visualization(visualizer, model, inputs):
    training_mode = model.training
    model.eval()

    outputs = visualizer.inference(model, inputs)
    visualizer.process(inputs, outputs)

    model.train(training_mode)
    
with EventStorage(start_iter) as storage: