ResizeShortestEdge augmentation makes boxes incorrect

[Shuntw6096]

Hi, I want to customize my data mapper with Albumentations, but when I trained faster rcnn with visualize_training

detectron2/detectron2/modeling/meta_arch/rcnn.py

Line 87 in a24729a

def visualize_training(self, batched_inputs, proposals):

)
I found GT annotation(box) are incorrect. And I checked box before and after ResizeShortestEdge, it seems that ResizeShortestEdge does not transform boxes, but I use default dataset mapper, the annotation on tensorboard (visualize_training) are correct, is there any function to solve this problem?

before(correct) image

after(wrong) image

class mymapper:
    def __init__(self, cfg):
            self.min_size = cfg.INPUT.MIN_SIZE_TRAIN
            self.max_size = cfg.INPUT.MAX_SIZE_TRAIN
            self.sample_style = cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING
    def __call__(self, dataset_dict):
        from detectron2.structures import BoxMode
        from pathlib import Path
        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")
        # See "Data Augmentation" tutorial for details usage
        transform = A.Compose([
        A.RandomScale(scale_limit=(-0.9, 1), p=1), #LargeScaleJitter from scale of 0.1 to 2
        A.PadIfNeeded(512, 832, border_mode=0), #constant 0 border
        A.RandomCrop(512, 832),
        A.HorizontalFlip(p=0.5),
        ], bbox_params=A.BboxParams(format="pascal_voc", )
        )
        anns = [ann['bbox'] + [ann['category_id']] for ann in dataset_dict['annotations']]
        tr_output = transform(image=image, bboxes=anns)
        image = tr_output['image']
        height, width, _ =image.shape
        anns = tr_output['bboxes']
        # visualize(image, [ann[:-1] for ann in anns], dataset_dict["file_name"])
        anns = [{'category_id':ann[-1], 'bbox':ann[:-1], 'bbox_mode':BoxMode.XYXY_ABS} for ann in anns]
        auginput = T.AugInput(image)
        transform = T.ResizeShortestEdge(self.min_size, self.max_size, self.sample_style)(auginput)
        image = torch.from_numpy(auginput.image.transpose(2, 0, 1))

        annos = [
            utils.transform_instance_annotations(annotation, [transform], image.shape[1:])
            for annotation in dataset_dict.pop("annotations")
        ]
        ins = utils.annotations_to_instances(annos, image.shape[1:])
        from detectron2.utils.visualizer import Visualizer
        img = utils.convert_image_to_rgb(image.permute(1, 2, 0), 'BGR')
        
        v_gt = Visualizer(img, None)
        v_gt = v_gt.overlay_instances(boxes=ins.gt_boxes)
        v_gt.save(str(OUTPUT_DIR/'imgs'/Path(dataset_dict["file_name"]).name))
        return {
        "file_name":dataset_dict["file_name"],
        "image_id": dataset_dict["image_id"],
        "height": height,
        "width": width,
        "image": image,
        "instances": utils.annotations_to_instances(annos, image.shape[1:])
        }
class Trainer(DefaultTrainer):
    """
    We use the "DefaultTrainer" which contains pre-defined default logic for
    standard training workflow. They may not work for you, especially if you
    are working on a new research project. In that case you can write your
    own training loop. You can use "tools/plain_train_net.py" as an example.
    """

    @classmethod
    def build_evaluator(cls, cfg, dataset_name, output_folder=None):
        return build_evaluator(cfg, dataset_name, output_folder)

    @classmethod
    def test_with_TTA(cls, cfg, model):
        logger = logging.getLogger("detectron2.trainer")
        # In the end of training, run an evaluation with TTA
        # Only support some R-CNN models.
        logger.info("Running inference with test-time augmentation ...")
        model = GeneralizedRCNNWithTTA(cfg, model)
        evaluators = [
            cls.build_evaluator(
                cfg, name, output_folder=os.path.join(cfg.OUTPUT_DIR, "inference_TTA")
            )
            for name in cfg.DATASETS.TEST
        ]
        res = cls.test(cfg, model, evaluators)
        res = OrderedDict({k + "_TTA": v for k, v in res.items()})
        return res

    @classmethod
    def build_train_loader(cls, cfg):
        """
        Returns:
            iterable

        It now calls :func:`detectron2.data.build_detection_train_loader`.
        Overwrite it if you'd like a different data loader.
        """
        return build_detection_train_loader(cfg, mapper=mymapper(cfg))

Thanks.

[Shuntw6096]

        annos = [
            utils.transform_instance_annotations(annotation, [transform], image.shape[1:])
            for annotation in dataset_dict.pop("annotations")
        ]

to

        annos = [
            utils.transform_instance_annotations(annotation, [transform], image.shape[1:])
            for annotation in anns
        ]

The code is copied from doc, I forget to check variable, and it works.