Compatibility for Instance Segmentation

[gunesevitan]

I'm not actually suggesting to do wbf or nms on masks but I think those methods should be compatible with masks for instance segmentation. I haven't tried any methods in this repository for instance segmentation yet but I think it makes sense to get more refined masks with wbf and nms.

I'm currently doing this by adding masks argument to prepare_boxes and nms_method functions.

def prepare_boxes(boxes, scores, labels, masks=None):
    result_boxes = boxes.copy()

    cond = (result_boxes < 0)
    cond_sum = cond.astype(np.int32).sum()
    if cond_sum > 0:
        print('Warning. Fixed {} boxes coordinates < 0'.format(cond_sum))
        result_boxes[cond] = 0

    cond = (result_boxes > 1)
    cond_sum = cond.astype(np.int32).sum()
    if cond_sum > 0:
        print('Warning. Fixed {} boxes coordinates > 1. Check that your boxes was normalized at [0, 1]'.format(cond_sum))
        result_boxes[cond] = 1

    boxes1 = result_boxes.copy()
    result_boxes[:, 0] = np.min(boxes1[:, [0, 2]], axis=1)
    result_boxes[:, 2] = np.max(boxes1[:, [0, 2]], axis=1)
    result_boxes[:, 1] = np.min(boxes1[:, [1, 3]], axis=1)
    result_boxes[:, 3] = np.max(boxes1[:, [1, 3]], axis=1)

    area = (result_boxes[:, 2] - result_boxes[:, 0]) * (result_boxes[:, 3] - result_boxes[:, 1])
    cond = (area == 0)
    cond_sum = cond.astype(np.int32).sum()
    if cond_sum > 0:
        print('Warning. Removed {} boxes with zero area!'.format(cond_sum))
        result_boxes = result_boxes[area > 0]
        scores = scores[area > 0]
        labels = labels[area > 0]
        if masks is not None:
            masks = masks[area > 0]

    return result_boxes, scores, labels, masks

def nms_method(boxes, scores, labels, masks=None, method=3, iou_thr=0.5, sigma=0.5, thresh=0.001, weights=None):
    """
    :param boxes: list of boxes predictions from each model, each box is 4 numbers. 
    It has 3 dimensions (models_number, model_preds, 4)
    Order of boxes: x1, y1, x2, y2. We expect float normalized coordinates [0; 1] 
    :param scores: list of scores for each model 
    :param labels: list of labels for each model
    :param method: 1 - linear soft-NMS, 2 - gaussian soft-NMS, 3 - standard NMS
    :param iou_thr: IoU value for boxes to be a match 
    :param sigma: Sigma value for SoftNMS
    :param thresh: threshold for boxes to keep (important for SoftNMS)
    :param weights: list of weights for each model. Default: None, which means weight == 1 for each model

    :return: boxes: boxes coordinates (Order of boxes: x1, y1, x2, y2). 
    :return: scores: confidence scores
    :return: labels: boxes labels
    """

    # If weights are specified
    if weights is not None:
        if len(boxes) != len(weights):
            print('Incorrect number of weights: {}. Must be: {}. Skip it'.format(len(weights), len(boxes)))
        else:
            weights = np.array(weights)
            for i in range(len(weights)):
                scores[i] = (np.array(scores[i]) * weights[i]) / weights.sum()

    # We concatenate everything
    boxes = np.concatenate(boxes)
    scores = np.concatenate(scores)
    labels = np.concatenate(labels)
    if masks is not None:
        masks = np.concatenate(masks)

    # Fix coordinates and removed zero area boxes
    boxes, scores, labels, masks = prepare_boxes(boxes, scores, labels, masks)

    # Run NMS independently for each label
    unique_labels = np.unique(labels)
    final_boxes = []
    final_scores = []
    final_labels = []
    if masks is not None:
        final_masks = []

    for l in unique_labels:
        condition = (labels == l)
        boxes_by_label = boxes[condition]
        scores_by_label = scores[condition]
        labels_by_label = np.array([l] * len(boxes_by_label))
        if masks is not None:
            masks_by_label = masks[condition]

        if method != 3:
            keep = cpu_soft_nms_float(boxes_by_label.copy(), scores_by_label.copy(), Nt=iou_thr, sigma=sigma, thresh=thresh, method=method)
        else:
            # Use faster function
            keep = nms_float_fast(boxes_by_label, scores_by_label, thresh=iou_thr)

        final_boxes.append(boxes_by_label[keep])
        final_scores.append(scores_by_label[keep])
        final_labels.append(labels_by_label[keep])
        if masks is not None:
            final_masks.append(masks_by_label[keep])
    final_boxes = np.concatenate(final_boxes)
    final_scores = np.concatenate(final_scores)
    final_labels = np.concatenate(final_labels)
    final_masks = np.concatenate(final_masks)

    return final_boxes, final_scores, final_labels, final_masks

def nms(boxes, scores, labels, masks=None, iou_thr=0.5, weights=None):
    """
    Short call for standard NMS 
    
    :param boxes: 
    :param scores: 
    :param labels: 
    :param iou_thr: 
    :param weights: 
    :return: 
    """
    return nms_method(boxes, scores, labels, masks, method=3, iou_thr=iou_thr, weights=weights)


def soft_nms(boxes, scores, labels, masks=None, method=2, iou_thr=0.5, sigma=0.5, thresh=0.001, weights=None):
    """
    Short call for Soft-NMS
     
    :param boxes: 
    :param scores: 
    :param labels: 
    :param method: 
    :param iou_thr: 
    :param sigma: 
    :param thresh: 
    :param weights: 
    :return: 
    """
    return nms_method(boxes, scores, labels, masks, method=method, iou_thr=iou_thr, sigma=sigma, thresh=thresh, weights=weights)

I know it looks like error prone with this fast work around but do you think this feature deserves a pull request? @ZFTurbo

[ZFTurbo]

We can add it, since it's common usage case for NMS.

But I already see the error in code. final_masks = np.concatenate(final_masks) It will fail.

And please add comment about parameter.