yolov8中obb旋转框的标签转换

yolo obb格式的数据集标注方法是多边形标注法，即标注出来旋转框的四个xy坐标，即是一个长度为8的张量，但在yolo的obb检测头中，并不是直接预测8个坐标值，而是预测xywhθ，这就牵涉到ground truth标签的转换 。

直接对yolov8代码进行调试，可以发现在trainer.py中，

self.train_loader = self.get_dataloader(self.trainset, batch_size=batch_size, rank=RANK, mode="train")

继续进入，最终会进入到dataset.py：

class YOLODataset(BaseDataset):
    """
    Dataset class for loading object detection and/or segmentation labels in YOLO format.

    Args:
        data (dict, optional): A dataset YAML dictionary. Defaults to None.
        task (str): An explicit arg to point current task, Defaults to 'detect'.

    Returns:
        (torch.utils.data.Dataset): A PyTorch dataset object that can be used for training an object detection model.
    """

    def __init__(self, *args, data=None, task="detect", **kwargs):
        """Initializes the YOLODataset with optional configurations for segments and keypoints."""
        self.use_segments = task == "segment"
        self.use_keypoints = task == "pose"
        self.use_obb = task == "obb"
        self.data = data
        assert not (self.use_segments and self.use_keypoints), "Can not use both segments and keypoints."
        super().__init__(*args, **kwargs)

    def cache_labels(self, path=Path("./labels.cache")):
        """
        Cache dataset labels, check images and read shapes.

        Args:
            path (Path): Path where to save the cache file. Default is Path('./labels.cache').

        Returns:
            (dict): labels.
        """
        x = {"labels": []}
        nm, nf, ne, nc, msgs = 0, 0, 0, 0, []  # number missing, found, empty, corrupt, messages
        desc = f"{self.prefix}Scanning {path.parent / path.stem}..."
        total = len(self.im_files)
        nkpt, ndim = self.data.get("kpt_shape", (0, 0))
        if self.use_keypoints and (nkpt <= 0 or ndim not in {2, 3}):
            raise ValueError(
                "'kpt_shape' in data.yaml missing or incorrect. Should be a list with [number of "
                "keypoints, number of dims (2 for x,y or 3 for x,y,visible)], i.e. 'kpt_shape: [17, 3]'"
            )
        with ThreadPool(NUM_THREADS) as pool:
            results = pool.imap(
                func=verify_image_label,
                iterable=zip(
                    self.im_files,
                    self.label_files,
                    repeat(self.prefix),
                    repeat(self.use_keypoints),
                    repeat(len(self.data["names"])),
                    repeat(nkpt),
                    repeat(ndim),
                ),
            )
            pbar = TQDM(results, desc=desc, total=total)
            for im_file, lb, shape, segments, keypoint, nm_f, nf_f, ne_f, nc_f, msg in pbar:
                nm += nm_f
                nf += nf_f
                ne += ne_f
                nc += nc_f
                if im_file:
                    x["labels"].append(
                        {
                            "im_file": im_file,
                            "shape": shape,
                            "cls": lb[:, 0:1],  # n, 1
                            "bboxes": lb[:, 1:],  # n, 4
                            "segments": segments,
                            "keypoints": keypoint,
                            "normalized": True,
                            "bbox_format": "xywh",
                        }
                    )
                if msg:
                    msgs.append(msg)
                pbar.desc = f"{desc} {nf} images, {nm + ne} backgrounds, {nc} corrupt"
            pbar.close()

        if msgs:
            LOGGER.info("\n".join(msgs))
        if nf == 0:
            LOGGER.warning(f"{self.prefix}WARNING ⚠️ No labels found in {path}. {HELP_URL}")
        x["hash"] = get_hash(self.label_files + self.im_files)
        x["results"] = nf, nm, ne, nc, len(self.im_files)
        x["msgs"] = msgs  # warnings
        save_dataset_cache_file(self.prefix, path, x, DATASET_CACHE_VERSION)
        return x

    def get_labels(self):
        """Returns dictionary of labels for YOLO training."""
        self.label_files = img2label_paths(self.im_files)
        cache_path = Path(self.label_files[0]).parent.with_suffix(".cache")
        try:
            cache, exists = load_dataset_cache_file(cache_path), True  # attempt to load a *.cache file
            assert cache["version"] == DATASET_CACHE_VERSION  # matches current version
            assert cache["hash"] == get_hash(self.label_files + self.im_files)  # identical hash
        except (FileNotFoundError, AssertionError, AttributeError):
            cache, exists = self.cache_labels(cache_path), False  # run cache ops

        # Display cache
        nf, nm, ne, nc, n = cache.pop("results")  # found, missing, empty, corrupt, total
        if exists and LOCAL_RANK in {-1, 0}:
            d = f"Scanning {cache_path}... {nf} images, {nm + ne} backgrounds, {nc} corrupt"
            TQDM(None, desc=self.prefix + d, total=n, initial=n)  # display results
            if cache["msgs"]:
                LOGGER.info("\n".join(cache["msgs"]))  # display warnings

        # Read cache
        [cache.pop(k) for k in ("hash", "version", "msgs")]  # remove items
        labels = cache["labels"]
        if not labels:
            LOGGER.warning(f"WARNING ⚠️ No images found in {cache_path}, training may not work correctly. {HELP_URL}")
        self.im_files = [lb["im_file"] for lb in labels]  # update im_files

        # Check if the dataset is all boxes or all segments
        lengths = ((len(lb["cls"]), len(lb["bboxes"]), len(lb["segments"])) for lb in labels)
        len_cls, len_boxes, len_segments = (sum(x) for x in zip(*lengths))
        if len_segments and len_boxes != len_segments:
            LOGGER.warning(
                f"WARNING ⚠️ Box and segment counts should be equal, but got len(segments) = {len_segments}, "
                f"len(boxes) = {len_boxes}. To resolve this only boxes will be used and all segments will be removed. "
                "To avoid this please supply either a detect or segment dataset, not a detect-segment mixed dataset."
            )
            for lb in labels:
                lb["segments"] = []
        if len_cls == 0:
            LOGGER.warning(f"WARNING ⚠️ No labels found in {cache_path}, training may not work correctly. {HELP_URL}")
        return labels

    def build_transforms(self, hyp=None):
        """Builds and appends transforms to the list."""
        if self.augment:
            hyp.mosaic = hyp.mosaic if self.augment and not self.rect else 0.0
            hyp.mixup = hyp.mixup if self.augment and not self.rect else 0.0
            transforms = v8_transforms(self, self.imgsz, hyp)
        else:
            transforms = Compose([LetterBox(new_shape=(self.imgsz, self.imgsz), scaleup=False)])
        transforms.append(
            Format(
                bbox_format="xywh",
                normalize=True,
                return_mask=self.use_segments,
                return_keypoint=self.use_keypoints,
                return_obb=self.use_obb,
                batch_idx=True,
                mask_ratio=hyp.mask_ratio,
                mask_overlap=hyp.overlap_mask,
                bgr=hyp.bgr if self.augment else 0.0,  # only affect training.
            )
        )
        return transforms

    def close_mosaic(self, hyp):
        """Sets mosaic, copy_paste and mixup options to 0.0 and builds transformations."""
        hyp.mosaic = 0.0  # set mosaic ratio=0.0
        hyp.copy_paste = 0.0  # keep the same behavior as previous v8 close-mosaic
        hyp.mixup = 0.0  # keep the same behavior as previous v8 close-mosaic
        self.transforms = self.build_transforms(hyp)

    def update_labels_info(self, label):
        """
        Custom your label format here.

        Note:
            cls is not with bboxes now, classification and semantic segmentation need an independent cls label
            Can also support classification and semantic segmentation by adding or removing dict keys there.
        """
        bboxes = label.pop("bboxes")
        segments = label.pop("segments", [])
        keypoints = label.pop("keypoints", None)
        bbox_format = label.pop("bbox_format")
        normalized = label.pop("normalized")

        # NOTE: do NOT resample oriented boxes
        segment_resamples = 100 if self.use_obb else 1000
        if len(segments) > 0:
            # list[np.array(1000, 2)] * num_samples
            # (N, 1000, 2)
            segments = np.stack(resample_segments(segments, n=segment_resamples), axis=0)
        else:
            segments = np.zeros((0, segment_resamples, 2), dtype=np.float32)
        label["instances"] = Instances(bboxes, segments, keypoints, bbox_format=bbox_format, normalized=normalized)
        return label

    @staticmethod
    def collate_fn(batch):
        """Collates data samples into batches."""
        new_batch = {}
        keys = batch[0].keys()
        values = list(zip(*[list(b.values()) for b in batch]))
        for i, k in enumerate(keys):
            value = values[i]
            if k == "img":
                value = torch.stack(value, 0)
            if k in {"masks", "keypoints", "bboxes", "cls", "segments", "obb"}:
                value = torch.cat(value, 0)
            new_batch[k] = value
        new_batch["batch_idx"] = list(new_batch["batch_idx"])
        for i in range(len(new_batch["batch_idx"])):
            new_batch["batch_idx"][i] += i  # add target image index for build_targets()
        new_batch["batch_idx"] = torch.cat(new_batch["batch_idx"], 0)
        return new_batch

其中最关键的部分是：

def cache_labels(self, path=Path("./labels.cache")):

这个方法会读取数据集，并生成cache文件，cache文件包含了标注的各种信息，这个方法最关键的代码是：

        with ThreadPool(NUM_THREADS) as pool:
            results = pool.imap(
                func=verify_image_label,
                iterable=zip(
                    self.im_files,
                    self.label_files,
                    repeat(self.prefix),
                    repeat(self.use_keypoints),
                    repeat(len(self.data["names"])),
                    repeat(nkpt),
                    repeat(ndim),
                ),
            )

可以看到，开启了一个多线程，调用了verify_image_label函数，所以这个函数非常关键，看它的代码：

ef verify_image_label(args):
    """Verify one image-label pair."""
    im_file, lb_file, prefix, keypoint, num_cls, nkpt, ndim = args
    # Number (missing, found, empty, corrupt), message, segments, keypoints
    nm, nf, ne, nc, msg, segments, keypoints = 0, 0, 0, 0, "", [], None
    try:
        # Verify images
        im = Image.open(im_file)
        im.verify()  # PIL verify
        shape = exif_size(im)  # image size
        shape = (shape[1], shape[0])  # hw
        assert (shape[0] > 9) & (shape[1] > 9), f"image size {shape} <10 pixels"
        assert im.format.lower() in IMG_FORMATS, f"invalid image format {im.format}. {FORMATS_HELP_MSG}"
        if im.format.lower() in {"jpg", "jpeg"}:
            with open(im_file, "rb") as f:
                f.seek(-2, 2)
                if f.read() != b"\xff\xd9":  # corrupt JPEG
                    ImageOps.exif_transpose(Image.open(im_file)).save(im_file, "JPEG", subsampling=0, quality=100)
                    msg = f"{prefix}WARNING ⚠️ {im_file}: corrupt JPEG restored and saved"

        # Verify labels
        if os.path.isfile(lb_file):
            nf = 1  # label found
            with open(lb_file) as f:
                lb = [x.split() for x in f.read().strip().splitlines() if len(x)]
                if any(len(x) > 6 for x in lb) and (not keypoint):  # is segment
                    classes = np.array([x[0] for x in lb], dtype=np.float32)
                    segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb]  # (cls, xy1...)
                    lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)
                lb = np.array(lb, dtype=np.float32)
            nl = len(lb)
            if nl:
                if keypoint:
                    assert lb.shape[1] == (5 + nkpt * ndim), f"labels require {(5 + nkpt * ndim)} columns each"
                    points = lb[:, 5:].reshape(-1, ndim)[:, :2]
                else:
                    assert lb.shape[1] == 5, f"labels require 5 columns, {lb.shape[1]} columns detected"
                    points = lb[:, 1:]
                assert points.max() <= 1, f"non-normalized or out of bounds coordinates {points[points > 1]}"
                assert lb.min() >= 0, f"negative label values {lb[lb < 0]}"

                # All labels
                max_cls = lb[:, 0].max()  # max label count
                assert max_cls <= num_cls, (
                    f"Label class {int(max_cls)} exceeds dataset class count {num_cls}. "
                    f"Possible class labels are 0-{num_cls - 1}"
                )
                _, i = np.unique(lb, axis=0, return_index=True)
                if len(i) < nl:  # duplicate row check
                    lb = lb[i]  # remove duplicates
                    if segments:
                        segments = [segments[x] for x in i]
                    msg = f"{prefix}WARNING ⚠️ {im_file}: {nl - len(i)} duplicate labels removed"
            else:
                ne = 1  # label empty
                lb = np.zeros((0, (5 + nkpt * ndim) if keypoint else 5), dtype=np.float32)
        else:
            nm = 1  # label missing
            lb = np.zeros((0, (5 + nkpt * ndim) if keypoints else 5), dtype=np.float32)
        if keypoint:
            keypoints = lb[:, 5:].reshape(-1, nkpt, ndim)
            if ndim == 2:
                kpt_mask = np.where((keypoints[..., 0] < 0) | (keypoints[..., 1] < 0), 0.0, 1.0).astype(np.float32)
                keypoints = np.concatenate([keypoints, kpt_mask[..., None]], axis=-1)  # (nl, nkpt, 3)
        lb = lb[:, :5]
        return im_file, lb, shape, segments, keypoints, nm, nf, ne, nc, msg
    except Exception as e:
        nc = 1
        msg = f"{prefix}WARNING ⚠️ {im_file}: ignoring corrupt image/label: {e}"
        return [None, None, None, None, None, nm, nf, ne, nc, msg]

直接看它的labels部分，

        if os.path.isfile(lb_file):
            nf = 1  # label found
            with open(lb_file) as f:
                lb = [x.split() for x in f.read().strip().splitlines() if len(x)]
                if any(len(x) > 6 for x in lb) and (not keypoint):  # is segment
                    classes = np.array([x[0] for x in lb], dtype=np.float32)
                    segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb]  # (cls, xy1...)
                    lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)
                lb = np.array(lb, dtype=np.float32)
            nl = len(lb)

可以看到，只要是标注的坐标个数大于6，就开启segments模式，请注意，这并不代表是进行图像分割，只是这里表示用了segment：

可以看到，lb变量每个子数据长度都是9，其中第一个是object类别，其余8个是多边形的顶点xy，

然后得到的segment则是很多个4*2的数组，这是对lb通过reshape得到的

lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)



def segments2boxes(segments):
    """
    It converts segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh)

    Args:
        segments (list): list of segments, each segment is a list of points, each point is a list of x, y coordinates

    Returns:
        (np.ndarray): the xywh coordinates of the bounding boxes.
    """
    boxes = []
    for s in segments:
        x, y = s.T  # segment xy
        boxes.append([x.min(), y.min(), x.max(), y.max()])  # cls, xyxy
    return xyxy2xywh(np.array(boxes))  # cls, xywh

这里我们可以知道，通过segments2boxes方法，得到旋转框的最大最小顶点 ，作为xyxy表示法的坐标，再转换为xywh

train.py:

    def get_dataloader(self, dataset_path, batch_size=16, rank=0, mode="train"):
        """Construct and return dataloader."""
        assert mode in {"train", "val"}, f"Mode must be 'train' or 'val', not {mode}."
        with torch_distributed_zero_first(rank):  # init dataset *.cache only once if DDP
            dataset = self.build_dataset(dataset_path, mode, batch_size)
        shuffle = mode == "train"
        if getattr(dataset, "rect", False) and shuffle:
            LOGGER.warning("WARNING ⚠️ 'rect=True' is incompatible with DataLoader shuffle, setting shuffle=False")
            shuffle = False
        workers = self.args.workers if mode == "train" else self.args.workers * 2
        return build_dataloader(dataset, batch_size, workers, shuffle, rank)  # return dataloader

这里的 dataset = self.build_dataset(dataset_path, mode, batch_size)

得到了：

其中labels中又包含了

可以看到obb信息被存储在了segment中

return build_dataloader(dataset, batch_size, workers, shuffle, rank)  # return dataloader

    def update_labels_info(self, label):
        """
        Custom your label format here.

        Note:
            cls is not with bboxes now, classification and semantic segmentation need an independent cls label
            Can also support classification and semantic segmentation by adding or removing dict keys there.
        """
        bboxes = label.pop("bboxes")
        segments = label.pop("segments", [])
        keypoints = label.pop("keypoints", None)
        bbox_format = label.pop("bbox_format")
        normalized = label.pop("normalized")

        # NOTE: do NOT resample oriented boxes
        segment_resamples = 100 if self.use_obb else 1000
        if len(segments) > 0:
            # list[np.array(1000, 2)] * num_samples
            # (N, 1000, 2)
            segments = np.stack(resample_segments(segments, n=segment_resamples), axis=0)
        else:
            segments = np.zeros((0, segment_resamples, 2), dtype=np.float32)
        label["instances"] = Instances(bboxes, segments, keypoints, bbox_format=bbox_format, normalized=normalized)
        return label

再回过头去看这个函数，可以发现，

segment_resamples = 100 if self.use_obb else 1000
        if len(segments) > 0:
            # list[np.array(1000, 2)] * num_samples
            # (N, 1000, 2)
            segments = np.stack(resample_segments(segments, n=segment_resamples), axis=0)
        else:
            segments = np.zeros((0, segment_resamples, 2), dtype=np.float32)

如果使用了obb，会对分割点进行重采样，也就是原本4*2就能表示一个真实框，现在上采样成100*2:

def resample_segments(segments, n=1000):
    """
    Inputs a list of segments (n,2) and returns a list of segments (n,2) up-sampled to n points each.

    Args:
        segments (list): a list of (n,2) arrays, where n is the number of points in the segment.
        n (int): number of points to resample the segment to. Defaults to 1000

    Returns:
        segments (list): the resampled segments.
    """
    for i, s in enumerate(segments):
        s = np.concatenate((s, s[0:1, :]), axis=0)
        x = np.linspace(0, len(s) - 1, n)
        xp = np.arange(len(s))
        segments[i] = (
            np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)], dtype=np.float32).reshape(2, -1).T
        )  # segment xy
    return segments

这样相当于用100个点包围了object，然后在数据增强transforms中：

    def build_transforms(self, hyp=None):
        """Builds and appends transforms to the list."""
        if self.augment:
            hyp.mosaic = hyp.mosaic if self.augment and not self.rect else 0.0
            hyp.mixup = hyp.mixup if self.augment and not self.rect else 0.0
            transforms = v8_transforms(self, self.imgsz, hyp)
        else:
            transforms = Compose([LetterBox(new_shape=(self.imgsz, self.imgsz), scaleup=False)])
        transforms.append(
            Format(
                bbox_format="xywh",
                normalize=True,
                return_mask=self.use_segments,
                return_keypoint=self.use_keypoints,
                return_obb=self.use_obb,
                batch_idx=True,
                mask_ratio=hyp.mask_ratio,
                mask_overlap=hyp.overlap_mask,
                bgr=hyp.bgr if self.augment else 0.0,  # only affect training.
            )
        )
        return transforms

 

class Format:
"""省略一些代码"""

    def __call__(self, labels):
        """Return formatted image, classes, bounding boxes & keypoints to be used by 'collate_fn'."""
        img = labels.pop("img")
        h, w = img.shape[:2]
        cls = labels.pop("cls")
        instances = labels.pop("instances")
        instances.convert_bbox(format=self.bbox_format)
        instances.denormalize(w, h)
        nl = len(instances)

        if self.return_mask:
            if nl:
                masks, instances, cls = self._format_segments(instances, cls, w, h)
                masks = torch.from_numpy(masks)
            else:
                masks = torch.zeros(
                    1 if self.mask_overlap else nl, img.shape[0] // self.mask_ratio, img.shape[1] // self.mask_ratio
                )
            labels["masks"] = masks
        labels["img"] = self._format_img(img)
        labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl)
        labels["bboxes"] = torch.from_numpy(instances.bboxes) if nl else torch.zeros((nl, 4))
        if self.return_keypoint:
            labels["keypoints"] = torch.from_numpy(instances.keypoints)
            if self.normalize:
                labels["keypoints"][..., 0] /= w
                labels["keypoints"][..., 1] /= h
        if self.return_obb:
            labels["bboxes"] = (
                xyxyxyxy2xywhr(torch.from_numpy(instances.segments)) if len(instances.segments) else torch.zeros((0, 5))
            )
        # NOTE: need to normalize obb in xywhr format for width-height consistency
        if self.normalize:
            labels["bboxes"][:, [0, 2]] /= w
            labels["bboxes"][:, [1, 3]] /= h
        # Then we can use collate_fn
        if self.batch_idx:
            labels["batch_idx"] = torch.zeros(nl)
        return labels

这里会将刚才的100*2表示的目标转换成xywhθ，从而与yolov8obb预测的格式一致：

        if self.return_obb:
            labels["bboxes"] = (
                xyxyxyxy2xywhr(torch.from_numpy(instances.segments)) if len(instances.segments) else torch.zeros((0, 5))


def xyxyxyxy2xywhr(x):
    """
    Convert batched Oriented Bounding Boxes (OBB) from [xy1, xy2, xy3, xy4] to [xywh, rotation]. Rotation values are
    expected in degrees from 0 to 90.

    Args:
        x (numpy.ndarray | torch.Tensor): Input box corners [xy1, xy2, xy3, xy4] of shape (n, 8).

    Returns:
        (numpy.ndarray | torch.Tensor): Converted data in [cx, cy, w, h, rotation] format of shape (n, 5).
    """
    is_torch = isinstance(x, torch.Tensor)
    points = x.cpu().numpy() if is_torch else x
    points = points.reshape(len(x), -1, 2)
    rboxes = []
    for pts in points:
        # NOTE: Use cv2.minAreaRect to get accurate xywhr,
        # especially some objects are cut off by augmentations in dataloader.
        (cx, cy), (w, h), angle = cv2.minAreaRect(pts)
        rboxes.append([cx, cy, w, h, angle / 180 * np.pi])
    return torch.tensor(rboxes, device=x.device, dtype=x.dtype) if is_torch else np.asarray(rboxes)