line_iou.py

"""
Copyright (c) 2022 Kunyang Zhou

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
import paddle
import paddle.nn as nn

from ..builder import LOSSES

def line_iou(pred, target, img_w, length=15, aligned=True):
    '''
    Calculate the line iou value between predictions and targets
    Args:
        pred: lane predictions, shape: (num_pred, 72)
        target: ground truth, shape: (num_target, 72)
        img_w: image width
        length: extended radius
        aligned: True for iou loss calculation, False for pair-wise ious in assign
    '''
    px1 = pred - length
    px2 = pred + length
    tx1 = target - length
    tx2 = target + length

    if aligned:
        invalid_mask = target
        ovr = paddle.minimum(px2, tx2) - paddle.maximum(px1, tx1)
        union = paddle.maximum(px2, tx2) - paddle.minimum(px1, tx1)
    else:
        num_pred = pred.shape[0]
        invalid_mask = target.tile([num_pred, 1, 1])

        ovr = (paddle.minimum(px2[:, None, :], tx2[None, ...]) - paddle.maximum(
            px1[:, None, :], tx1[None, ...]))
        union = (paddle.maximum(px2[:, None, :], tx2[None, ...]) -
                 paddle.minimum(px1[:, None, :], tx1[None, ...]))

    invalid_masks = (invalid_mask < 0) | (invalid_mask >= img_w)

    ovr[invalid_masks] = 0.
    union[invalid_masks] = 0.
    iou = ovr.sum(axis=-1) / (union.sum(axis=-1) + 1e-9)
    return iou

def liou_loss(pred, target, img_w, length=15):
    return (1 - line_iou(pred, target, img_w, length)).mean()

def lane_iou(pred, 
             target, 
             img_h = 320, 
             img_w = 1640, 
             length=7.5, 
             max_dx=1e4,
             aligned=True):
        """
        Implementation of LaneIOU in CLRerNet
        Calculate the LaneIoU value between predictions and targets
        Args:
            pred: lane predictions, shape: (Nl, Nr), relative coordinate.
            target: ground truth, shape: (Nl, Nr), relative coordinate.
            img_h (int): original image height corresponding to the feature map region. 320 is the height of the image on CULane.
            img_w (int): original image width corresponding to the feature map region. 1640 is the width of the image on CULane.
        Returns:
            torch.Tensor: virtual lane half-widths for prediction at pre-defined rows, shape (Nl, Nr).
            torch.Tensor: virtual lane half-widths for GT at pre-defined rows, shape (Nl, Nr).
        Nl: number of lanes, Nr: number of rows.
        """    
        
        n_strips = pred.shape[1] - 1
        dy = img_h / n_strips * 2  # two horizontal grids
        _pred = pred.clone().detach()
        pred_dx = (
            _pred[:, 2:] - _pred[:, :-2]
        ) * img_w  # pred x difference across two horizontal grids
        pred_width = length * paddle.sqrt(pred_dx.pow(2) + dy**2) / dy
        pred_width = paddle.concat(
            [pred_width[:, 0:1], pred_width, pred_width[:, -1:]], axis=1
        )
        target_dx = (target[:, 2:] - target[:, :-2]) * img_w
        target_dx[paddle.abs(target_dx) > max_dx] = 0
        target_width = length * paddle.sqrt(target_dx.pow(2) + dy**2) / dy
        target_width = paddle.concat(
            [target_width[:, 0:1], target_width, target_width[:, -1:]], dim=1
        )

        px1 = pred - pred_width
        px2 = pred + pred_width
        tx1 = target - target_width
        tx2 = target + target_width

        if aligned:
            invalid_mask = target
            ovr = paddle.minimum(px2, tx2) - paddle.maximum(px1, tx1)
            union = paddle.maximum(px2, tx2) - paddle.minimum(px1, tx1)
        else:
            num_pred = pred.shape[0]
            invalid_mask = target.tile([num_pred, 1, 1])

            ovr = (paddle.minimum(px2[:, None, :], tx2[None, ...]) - paddle.maximum(
                px1[:, None, :], tx1[None, ...]))
            union = (paddle.maximum(px2[:, None, :], tx2[None, ...]) -
                    paddle.minimum(px1[:, None, :], tx1[None, ...]))

        invalid_masks = (invalid_mask < 0) | (invalid_mask >= img_w)

        ovr[invalid_masks] = 0.
        union[invalid_masks] = 0.
        iou = ovr.sum(axis=-1) / (union.sum(axis=-1) + 1e-9)
        return iou
        
def laneliou_loss(pred, target, img_w, length=7.5, **kwargs):
    return (1 - lane_iou(pred, target, img_w, length, **kwargs)).mean()

@LOSSES.register()
class Liou_loss(nn.Layer):
    """
    Line IoU loss in CLRNet
    """
    def __init__(self,cfg):
        super().__init__()
    
    def forward(self,pred, target, img_w, length=15):
        return liou_loss(pred, target, img_w, length)

@LOSSES.register()
class LaneIoU_loss(nn.Layer):
    """
    Lane IoU loss in CLRerNet
    """
    def __init__(self,cfg):
        super().__init__() 
           
    def forward(self,
                pred, 
                target, 
                img_w, 
                length=15,
                **kwargs):
        return laneliou_loss(pred, target, img_w, length, **kwargs)