Issue in Multitask Segmentation & Keypoint detection with Detectron2

[nicojany]

Hey there!

I know there are threads regarding training a segmentation/keypoint model simultaneously. However, I didn't find a final answer yet if this is possible in general or if I simply can stopp trying it :)...

I'm trying to use the pre-trained model: mask_rcnn_X_101_32x8d_FPN_3x for identifying roofs/ridges/obstacles simultaneously. I want to use segmentation annotation (which I'm using for polygon shapes --> roofs/obstacles/some others) and keypoint annotation (which I'm using for line shapes --> ridge/valleys) simultaneously.

My annotation datasets for training and validation consequently have segmentation annotations and keypoint annotations. I already added dummy arrays/values for segmentation annotations (keypoints array and num_keypoints = 2) to have a similar data structure within the annotations. However this didn't solve the problem. My annotations look like this:

Segmentation

        {
            "area": 160,
            "bbox": [
                145,
                337,
                16,
                11
            ],
            "category_id": 1,
            "id": 9,
            "image_id": 1,
            "iscrowd": 0,
            "segmentation": [
                [
                    145,
                    337,
                    145,
                    347,
                    160,
                    346,
                    160,
                    337
                ]
            ],
            "keypoints": [
                0,
                0,
                0,
                0,
                0,
                0
            ],
            "num_keypoints": 2
        }...

Keypoints:

        {
            "area": 3240,
            "bbox": [
                81,
                280,
                60,
                53
            ],
            "category_id": 4,
            "id": 1,
            "image_id": 1,
            "iscrowd": 0,
            "keypoints": [
                139,
                280,
                2,
                81,
                333,
                2
            ],
            "num_keypoints": 2,
            "segmentation": []
        }...

In my training file I activated model_mask and model_keypoint. My config looks like this:

def setup_cfg():
    cfg = get_cfg()
    cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml"))
    
    cfg.MODEL.MASK_ON = True
    cfg.MODEL.KEYPOINT_ON = True

    cfg.MODEL.DEVICE = "cpu"
    
    cfg.DATASETS.TRAIN = ("roof_train",)
    cfg.DATASETS.TEST = ("roof_val",)
    cfg.DATALOADER.NUM_WORKERS = 2
    cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml")
    cfg.MODEL.ROI_HEADS.NUM_CLASSES = 7
    cfg.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = 2
    cfg.SOLVER.IMS_PER_BATCH = 2
    cfg.SOLVER.BASE_LR = 0.001
    cfg.SOLVER.MAX_ITER = 500
    cfg.SOLVER.STEPS = [] 
    cfg.OUTPUT_DIR = ".../models/new"
    
    return cfg

Unfortunately it is not working. Errors:

AttributeError: Cannot find field 'gt_masks' in the given Instances!
or
ValueError: expected sequence of length 2 at dim 1 (got 0)
or
ValueError: Keypoint data has 2 points, but metadata contains 1 points!

with a custom_dataset_mapper:

class CustomDatasetMapper:
    def __init__(self, cfg, is_train=True):
        self.is_train = is_train
        self.augmentations = T.AugmentationList(utils.build_augmentation(cfg, is_train))
        self.img_format = cfg.INPUT.FORMAT
        self.keypoint_hflip_indices = (
            MetadataCatalog.get(cfg.DATASETS.TRAIN[0]).keypoint_flip_map
            if cfg.MODEL.KEYPOINT_ON and len(cfg.DATASETS.TRAIN)
            else None
        )

    def __call__(self, dataset_dict):
        dataset_dict = dataset_dict.copy()
        image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
        utils.check_image_size(dataset_dict, image)
        
        aug_input = T.AugInput(image)
        transforms = self.augmentations(aug_input)
        image = aug_input.image
        
        image_shape = image.shape[:2]
        dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))

        annos = [
            utils.transform_instance_annotations(obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices)
            for obj in dataset_dict.pop("annotations")
            if obj.get("iscrowd", 0) == 0
        ]

        instances = utils.annotations_to_instances(annos, image_shape)

        if not hasattr(instances, "gt_masks"):
            instances.gt_masks = torch.zeros((len(instances), *image_shape), dtype=torch.uint8)

        dataset_dict["instances"] = instances
        return dataset_dict

When only exporting the segmentation annotations and deactivating the model.keypoints or vice versa (exporting keypoint annotations only and deactivating model.mask) it works. However, multitask it seems not working...

Someone any idea on that?

Thanks in advance!

[solosophi]

Have you generated mask files(.png format) corresponding to the segmentation in your JSON file? If not you have to generate it and you can save it in the same directory as jpg image root.