Reference for ultralytics/utils/nms.py#

class TorchNMS:
    """Ultralytics custom NMS implementation optimized for YOLO.

    This class provides static methods for performing non-maximum suppression (NMS) operations on bounding boxes,
    including standard NMS and fast NMS.

    Methods:
        fast_nms: Fast-NMS using upper triangular matrix operations.
        nms: Optimized NMS with early termination that matches torchvision behavior exactly.

    Examples:
        Perform standard NMS on boxes and scores
        >>> boxes = torch.tensor([[0, 0, 10, 10], [5, 5, 15, 15]])
        >>> scores = torch.tensor([0.9, 0.8])
        >>> keep = TorchNMS.nms(boxes, scores, 0.5)
    """

@staticmethod
def fast_nms(
    boxes: torch.Tensor,
    scores: torch.Tensor,
    iou_threshold: float,
    use_triu: bool = True,
    iou_func=box_iou,
    exit_early: bool = True,
) -> torch.Tensor:
    """Fast-NMS implementation from https://arxiv.org/pdf/1904.02689 using upper triangular matrix operations.

    Args:
        boxes (torch.Tensor): Bounding boxes with shape (N, 4) in xyxy format, or (N, 5) in xywhr format when
            iou_func is batch_probiou.
        scores (torch.Tensor): Confidence scores with shape (N,).
        iou_threshold (float): IoU threshold for suppression.
        use_triu (bool): Whether to use torch.triu operator for upper triangular matrix operations. If False, an
            export-friendly path zeroes suppressed scores in place and returns all N indices sorted by score.
        iou_func (callable): Function to compute IoU between boxes.
        exit_early (bool): Whether to return early if there are no boxes.

    Returns:
        (torch.Tensor): Indices of boxes to keep after NMS, sorted by descending score.

    Examples:
        Apply NMS to a set of boxes
        >>> boxes = torch.tensor([[0, 0, 10, 10], [5, 5, 15, 15]])
        >>> scores = torch.tensor([0.9, 0.8])
        >>> keep = TorchNMS.fast_nms(boxes, scores, 0.5)
    """
    if boxes.numel() == 0 and exit_early:
        return torch.empty((0,), dtype=torch.int64, device=boxes.device)

    sorted_idx = torch.argsort(scores, descending=True)
    boxes = boxes[sorted_idx]
    ious = iou_func(boxes, boxes)
    if use_triu:
        ious = ious.triu_(diagonal=1)
        # NOTE: no if-else on len(boxes), keeping this path exportable
        pick = torch.nonzero((ious > iou_threshold).sum(0) <= 0).squeeze_(-1)
    else:
        n = boxes.shape[0]
        row_idx = torch.arange(n, device=boxes.device).view(-1, 1).expand(-1, n)
        col_idx = torch.arange(n, device=boxes.device).view(1, -1).expand(n, -1)
        upper_mask = row_idx < col_idx
        ious = ious * upper_mask
        # Zeroing these scores ensures the additional indices would not affect the final results
        scores_ = scores[sorted_idx]
        scores_[~((ious > iou_threshold).sum(0) <= 0)] = 0
        scores[sorted_idx] = scores_  # update original tensor for NMSModel
        # NOTE: return indices with fixed length to avoid TFLite reshape error
        pick = torch.topk(scores_, scores_.shape[0]).indices
    return sorted_idx[pick]

@staticmethod
def nms(boxes: torch.Tensor, scores: torch.Tensor, iou_threshold: float) -> torch.Tensor:
    """Optimized NMS with early termination that matches torchvision behavior exactly.

    Args:
        boxes (torch.Tensor): Bounding boxes with shape (N, 4) in xyxy format.
        scores (torch.Tensor): Confidence scores with shape (N,).
        iou_threshold (float): IoU threshold for suppression.

    Returns:
        (torch.Tensor): Indices of boxes to keep after NMS.

    Examples:
        Apply NMS to a set of boxes
        >>> boxes = torch.tensor([[0, 0, 10, 10], [5, 5, 15, 15]])
        >>> scores = torch.tensor([0.9, 0.8])
        >>> keep = TorchNMS.nms(boxes, scores, 0.5)
    """
    if boxes.numel() == 0:
        return torch.empty((0,), dtype=torch.int64, device=boxes.device)

    # Pre-allocate and extract coordinates once
    x1, y1, x2, y2 = boxes.unbind(1)
    areas = (x2 - x1) * (y2 - y1)

    # Sort by scores descending
    order = scores.argsort(0, descending=True)

    # Pre-allocate keep list with maximum possible size
    keep = torch.zeros(order.numel(), dtype=torch.int64, device=boxes.device)
    keep_idx = 0
    while order.numel() > 0:
        i = order[0]
        keep[keep_idx] = i
        keep_idx += 1

        if order.numel() == 1:
            break
        # Vectorized IoU calculation for remaining boxes
        rest = order[1:]
        xx1 = torch.maximum(x1[i], x1[rest])
        yy1 = torch.maximum(y1[i], y1[rest])
        xx2 = torch.minimum(x2[i], x2[rest])
        yy2 = torch.minimum(y2[i], y2[rest])

        # Fast intersection and IoU
        w = (xx2 - xx1).clamp_(min=0)
        h = (yy2 - yy1).clamp_(min=0)
        inter = w * h
        # Early exit: skip IoU calculation if no intersection
        if inter.sum() == 0:
            # No overlaps with current box, keep all remaining boxes
            order = rest
            continue
        iou = inter / (areas[i] + areas[rest] - inter)
        # Keep boxes with IoU <= threshold
        order = rest[iou <= iou_threshold]

    return keep[:keep_idx]

def non_max_suppression(
    prediction,
    conf_thres: float = 0.25,
    iou_thres: float = 0.45,
    classes=None,
    agnostic: bool = False,
    multi_label: bool = False,
    max_det: int = 300,
    nc: int = 0,  # number of classes (optional)
    max_time_img: float = 0.05,
    max_nms: int = 30000,
    max_wh: int = 7680,
    rotated: bool = False,
    end2end: bool = False,
    return_idxs: bool = False,
):
    """Perform non-maximum suppression (NMS) on prediction results.

    Applies NMS to filter overlapping bounding boxes based on confidence and IoU thresholds. Supports multiple detection
    formats including standard boxes, rotated boxes, and masks. End-to-end outputs are only filtered by confidence,
    class, and max_det.

    Args:
        prediction (torch.Tensor | list | tuple): Predictions with shape (batch_size, 4 + num_classes + num_extra,
            num_boxes) containing xywh boxes, class scores, and optional extra channels (mask coefficients, keypoints,
            or angle), or end-to-end predictions with shape (batch_size, num_boxes, 6 + num_extra). For a list or tuple,
            the first element is used.
        conf_thres (float): Confidence threshold for filtering detections. Valid values are between 0.0 and 1.0.
        iou_thres (float): IoU threshold for NMS filtering. Valid values are between 0.0 and 1.0. The 0.45 default is
            not the cfg `iou` default (0.7), which predictors and validators always pass explicitly.
        classes (list[int], optional): List of class indices to consider. If None, all classes are considered.
        agnostic (bool): Whether to perform class-agnostic NMS.
        multi_label (bool): Whether each box can have multiple labels.
        max_det (int): Maximum number of detections to keep per image.
        nc (int): Number of classes; channels after 4 + nc are treated as extra channels. If 0, inferred as
            prediction.shape[1] - 4.
        max_time_img (float): Maximum time in seconds for processing one image.
        max_nms (int): Maximum number of boxes for NMS.
        max_wh (int): Per-class coordinate offset in pixels used to separate classes for class-aware NMS.
        rotated (bool): Whether to handle Oriented Bounding Boxes (OBB).
        end2end (bool): Whether the model is end-to-end and doesn't require NMS.
        return_idxs (bool): Whether to return the indices of kept detections.

    Returns:
        (list[torch.Tensor] | tuple[list[torch.Tensor], list[torch.Tensor]]): List of detections per image with shape
            (num_boxes, 6 + num_extra) containing (x1, y1, x2, y2, confidence, class, extra1, extra2, ...); with
            rotated=True the boxes stay in (x, y, w, h) format and the angle is the last extra channel. If
            return_idxs=True, returns a tuple of (output, keepi) where keepi contains, per image, the indices of the
            kept detections in the input predictions.
    """
    # Checks
    assert 0 <= conf_thres <= 1, f"Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0"
    assert 0 <= iou_thres <= 1, f"Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0"
    if isinstance(prediction, (list, tuple)):  # model in validation mode, output = (inference_out, loss_out)
        prediction = prediction[0]  # select only inference output
    if classes is not None:
        classes = torch.tensor(classes, device=prediction.device)

    if prediction.shape[-1] == 6 or end2end:  # end-to-end model (BNC, i.e. 1,300,6)
        output, keepi = [], []
        for pred in prediction:
            mask = pred[:, 4] > conf_thres
            if classes is not None:
                mask &= (pred[:, 5:6] == classes).any(1)
            idx = mask.nonzero(as_tuple=False).view(-1)[:max_det]
            output.append(pred[idx].float())  # half backend outputs: match the float32 rows the NMS branch returns
            keepi.append(idx)
        return (output, keepi) if return_idxs else output

    bs = prediction.shape[0]  # batch size (BCN, i.e. 1,84,6300)
    nc = nc or (prediction.shape[1] - 4)  # number of classes
    extra = prediction.shape[1] - nc - 4  # number of extra info
    mi = 4 + nc  # mask start index
    xc = prediction[:, 4:mi].amax(1) > conf_thres  # candidates
    xinds = torch.arange(prediction.shape[-1], device=prediction.device).expand(bs, -1)[..., None]  # to track idxs

    # Settings
    # min_wh = 2  # (pixels) minimum box width and height
    time_limit = 2.0 + max_time_img * bs  # seconds to quit after
    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)

    prediction = prediction.transpose(-1, -2)  # shape(1,84,6300) to shape(1,6300,84)

    t = time.time()
    output = [torch.zeros((0, 6 + extra), device=prediction.device)] * bs
    keepi = [torch.zeros((0, 1), device=prediction.device)] * bs  # to store the kept idxs
    use_torchvision = prediction.device.type != "xpu" and "torchvision" in sys.modules
    for xi, (x, xk) in enumerate(zip(prediction, xinds)):  # image index, (preds, preds indices)
        # Apply constraints
        # x[((x[:, 2:4] < min_wh) | (x[:, 2:4] > max_wh)).any(1), 4] = 0  # width-height
        filt = xc[xi]  # confidence
        x = x[filt]
        if not rotated:
            x[:, :4] = xywh2xyxy(x[:, :4])
        if return_idxs:
            xk = xk[filt]

        # If none remain process next image
        if not x.shape[0]:
            continue

        # Detections matrix nx6 (xyxy, conf, cls)
        box, cls, mask = x.split((4, nc, extra), 1)

        if multi_label:
            i, j = torch.where(cls > conf_thres)
            x = torch.cat((box[i], x[i, 4 + j, None], j[:, None].float(), mask[i]), 1)
            if return_idxs:
                xk = xk[i]
        else:  # best class only
            conf, j = cls.max(1, keepdim=True)
            filt = conf.view(-1) > conf_thres
            x = torch.cat((box, conf, j.float(), mask), 1)[filt]
            if return_idxs:
                xk = xk[filt]

        # Filter by class
        if classes is not None:
            filt = (x[:, 5:6] == classes).any(1)
            x = x[filt]
            if return_idxs:
                xk = xk[filt]

        # Check shape
        n = x.shape[0]  # number of boxes
        if not n:  # no boxes
            continue
        if n > max_nms:  # excess boxes
            filt = x[:, 4].argsort(descending=True)[:max_nms]  # sort by confidence and remove excess boxes
            x = x[filt]
            if return_idxs:
                xk = xk[filt]

        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
        scores = x[:, 4]  # scores
        if rotated:
            boxes = torch.cat((x[:, :2] + c, x[:, 2:4], x[:, -1:]), dim=-1)  # xywhr
            i = TorchNMS.fast_nms(boxes, scores, iou_thres, iou_func=batch_probiou)
        else:
            boxes = x[:, :4] + c  # boxes (offset by class)
            # Use torchvision if already imported; it has no XPU NMS kernel, but torch_npu CPU-falls back on NPU.
            if use_torchvision:
                import torchvision  # scope as slow import

                i = torchvision.ops.nms(boxes, scores, iou_thres)
            else:
                i = TorchNMS.nms(boxes, scores, iou_thres)
        i = i[:max_det]  # limit detections

        output[xi] = x[i]
        if return_idxs:
            keepi[xi] = xk[i].view(-1)
        if (time.time() - t) > time_limit:
            LOGGER.warning(f"NMS time limit {time_limit:.3f}s exceeded")
            break  # time limit exceeded

    return (output, keepi) if return_idxs else output