下载wordpress程序/全国推广优化网站

forward函数

 def forward(self, batch_dict):"""Args:batch_dict:batch_size:point_features: (N1 + N2 + N3 + ..., C) or (B, N, C)point_features_before_fusion: (N1 + N2 + N3 + ..., C)point_coords: (N1 + N2 + N3 + ..., 4) [bs_idx, x, y, z]point_labels (optional): (N1 + N2 + N3 + ...)gt_boxes (optional): (B, M, 8)Returns:batch_dict:point_cls_scores: (N1 + N2 + N3 + ..., 1)point_part_offset: (N1 + N2 + N3 + ..., 3)"""if self.model_cfg.get('USE_POINT_FEATURES_BEFORE_FUSION', False):point_features = batch_dict['point_features_before_fusion']else:point_features = batch_dict['point_features']#通过全连接层128-->256-->256-->3生成类别信息point_cls_preds = self.cls_layers(point_features)  # (total_points, num_class)#通过全连接层128-->256-->256-->8生成回归框信息point_box_preds = self.box_layers(point_features)  # (total_points, box_code_size)#在预测的3个类别中求出最大可能的类别作为标签信息，并经过sigmod函数point_cls_preds_max, _ = point_cls_preds.max(dim=-1)batch_dict['point_cls_scores'] = torch.sigmoid(point_cls_preds_max)ret_dict = {'point_cls_preds': point_cls_preds,'point_box_preds': point_box_preds}if self.training:#主要是生成每个点对应的真实的标签信息#以及真实框G相对于预测G_hat的框的参数偏移，每个点对应是1*8维向量targets_dict = self.assign_targets(batch_dict)ret_dict['point_cls_labels'] = targets_dict['point_cls_labels']ret_dict['point_box_labels'] = targets_dict['point_box_labels']if not self.training or self.predict_boxes_when_training:#求出每个点对应的预测的标签信息#以及P相对于预测的框G_hat的参数偏移，每个点对应是1*8维向量point_cls_preds, point_box_preds = self.generate_predicted_boxes(points=batch_dict['point_coords'][:, 1:4],point_cls_preds=point_cls_preds, point_box_preds=point_box_preds)batch_dict['batch_cls_preds'] = point_cls_predsbatch_dict['batch_box_preds'] = point_box_predsbatch_dict['batch_index'] = batch_dict['point_coords'][:, 0]batch_dict['cls_preds_normalized'] = Falseself.forward_ret_dict = ret_dictreturn batch_dict

注意：对于每一个point，point_box_preds是1×8维向量，8维分别表示[xt, yt, zt, dxt, dyt, dzt, cost, sint]，[xt, yt, zt]为中心点偏移量，[dxt, dyt, dzt]为长宽高偏移量，[cost, sint]为角度偏移量。

forward函数得到了每个前景点对应的真实标签值以及标注框信息；（self.assign_targets--------->self.assign_stack_targets-----> self.box_coder.encode_torch调用了PointResidualCoder类中的encode_torch函数）

得到了从G_hat到G的1*8维参数

---

每个前景点对应的预测标签值以及预测框信息；（self.generate_predicted_boxes--------->self.box_coder.decode_torch调用了PointResidualCoder类中的decode_torch函数）

得到了从P到G_hat的1*8维参数

得到这两组参数后用于后续计算损失时计算的box损失，采用的是L1回归损失

 point_loss_box_src = F.smooth_l1_loss(point_box_preds[None, ...], point_box_labels[None, ...], weights=reg_weights[None, ...])

边框回归(Bounding Box Regression)详解

PointResidualCoder

class PointResidualCoder(object):def __init__(self, code_size=8, use_mean_size=True, **kwargs):super().__init__()self.code_size = code_sizeself.use_mean_size = use_mean_sizeif self.use_mean_size:self.mean_size = torch.from_numpy(np.array(kwargs['mean_size'])).cuda().float()assert self.mean_size.min() > 0def encode_torch(self, gt_boxes, points, gt_classes=None):"""Args:gt_boxes: (N, 7 + C) [x, y, z, dx, dy, dz, heading, ...]points: (N, 3) [x, y, z]gt_classes: (N) [1, num_classes]Returns:box_coding: (N, 8 + C)"""gt_boxes[:, 3:6] = torch.clamp_min(gt_boxes[:, 3:6], min=1e-5)xg, yg, zg, dxg, dyg, dzg, rg, *cgs = torch.split(gt_boxes, 1, dim=-1)xa, ya, za = torch.split(points, 1, dim=-1)if self.use_mean_size:assert gt_classes.max() <= self.mean_size.shape[0]point_anchor_size = self.mean_size[gt_classes - 1]dxa, dya, dza = torch.split(point_anchor_size, 1, dim=-1)diagonal = torch.sqrt(dxa ** 2 + dya ** 2)xt = (xg - xa) / diagonalyt = (yg - ya) / diagonalzt = (zg - za) / dzadxt = torch.log(dxg / dxa)dyt = torch.log(dyg / dya)dzt = torch.log(dzg / dza)else:xt = (xg - xa)yt = (yg - ya)zt = (zg - za)dxt = torch.log(dxg)dyt = torch.log(dyg)dzt = torch.log(dzg)cts = [g for g in cgs]return torch.cat([xt, yt, zt, dxt, dyt, dzt, torch.cos(rg), torch.sin(rg), *cts], dim=-1)def decode_torch(self, box_encodings, points, pred_classes=None):"""Args:box_encodings: (N, 8 + C) [x, y, z, dx, dy, dz, cos, sin, ...]points: [x, y, z]pred_classes: (N) [1, num_classes]Returns:"""xt, yt, zt, dxt, dyt, dzt, cost, sint, *cts = torch.split(box_encodings, 1, dim=-1)xa, ya, za = torch.split(points, 1, dim=-1)if self.use_mean_size:assert pred_classes.max() <= self.mean_size.shape[0]point_anchor_size = self.mean_size[pred_classes - 1]dxa, dya, dza = torch.split(point_anchor_size, 1, dim=-1)diagonal = torch.sqrt(dxa ** 2 + dya ** 2)xg = xt * diagonal + xayg = yt * diagonal + yazg = zt * dza + zadxg = torch.exp(dxt) * dxadyg = torch.exp(dyt) * dyadzg = torch.exp(dzt) * dzaelse:xg = xt + xayg = yt + yazg = zt + zadxg, dyg, dzg = torch.split(torch.exp(box_encodings[..., 3:6]), 1, dim=-1)rg = torch.atan2(sint, cost)cgs = [t for t in cts]return torch.cat([xg, yg, zg, dxg, dyg, dzg, rg, *cgs], dim=-1)

decode_torch：如何通过point_box_preds的8维向量得到proposal的7维坐标？将每一个point原始xyz坐标加上坐标偏移量[xt, yt, zt]即可得到proposal中心点坐标，利用作者预设的point_anchor_size乘上长宽高偏移量[dxt, dyt, dzt]得到proposal长宽高，利用atan2函数计算角度heading。

论文出处
3D Object Detection for Autonomous Driving: A Review and New Outlooks

个人的理解是觉得这样可以同时优化生成的anchor大小并且可以调节中心坐标的偏移。

assign_targets

 def assign_targets(self, input_dict):"""Args:input_dict:point_features: (N1 + N2 + N3 + ..., C)batch_size:point_coords: (N1 + N2 + N3 + ..., 4) [bs_idx, x, y, z]gt_boxes (optional): (B, M, 8)Returns:point_cls_labels: (N1 + N2 + N3 + ...), long type, 0:background, -1:ignoredpoint_part_labels: (N1 + N2 + N3 + ..., 3)"""point_coords = input_dict['point_coords']gt_boxes = input_dict['gt_boxes']assert gt_boxes.shape.__len__() == 3, 'gt_boxes.shape=%s' % str(gt_boxes.shape)assert point_coords.shape.__len__() in [2], 'points.shape=%s' % str(point_coords.shape)batch_size = gt_boxes.shape[0]extend_gt_boxes = box_utils.enlarge_box3d(gt_boxes.view(-1, gt_boxes.shape[-1]), extra_width=self.model_cfg.TARGET_CONFIG.GT_EXTRA_WIDTH).view(batch_size, -1, gt_boxes.shape[-1])targets_dict = self.assign_stack_targets(points=point_coords, gt_boxes=gt_boxes, extend_gt_boxes=extend_gt_boxes,set_ignore_flag=True, use_ball_constraint=False,ret_part_labels=False, ret_box_labels=True)return targets_dict

extend_gt_boxes 主要是将groud truth boxex在长、宽、高方向上扩展

assign_stack_targets

#此函数传入的都是对应点的真实预测值和真实标注框def assign_stack_targets(self, points, gt_boxes, extend_gt_boxes=None,ret_box_labels=False, ret_part_labels=False,set_ignore_flag=True, use_ball_constraint=False, central_radius=2.0):"""Args:points: (N1 + N2 + N3 + ..., 4) [bs_idx, x, y, z]gt_boxes: (B, M, 8)extend_gt_boxes: [B, M, 8]ret_box_labels:ret_part_labels:set_ignore_flag:use_ball_constraint:central_radius:Returns:point_cls_labels: (N1 + N2 + N3 + ...), long type, 0:background, -1:ignoredpoint_box_labels: (N1 + N2 + N3 + ..., code_size)"""assert len(points.shape) == 2 and points.shape[1] == 4, 'points.shape=%s' % str(points.shape)assert len(gt_boxes.shape) == 3 and gt_boxes.shape[2] == 8, 'gt_boxes.shape=%s' % str(gt_boxes.shape)assert extend_gt_boxes is None or len(extend_gt_boxes.shape) == 3 and extend_gt_boxes.shape[2] == 8, \'extend_gt_boxes.shape=%s' % str(extend_gt_boxes.shape)assert set_ignore_flag != use_ball_constraint, 'Choose one only!'#将数据分批次处理batch_size = gt_boxes.shape[0]bs_idx = points[:, 0]point_cls_labels = points.new_zeros(points.shape[0]).long()point_box_labels = gt_boxes.new_zeros((points.shape[0], 8)) if ret_box_labels else Nonepoint_part_labels = gt_boxes.new_zeros((points.shape[0], 3)) if ret_part_labels else None#将数据分批次处理for k in range(batch_size):bs_mask = (bs_idx == k)#这里以*_single应该是中间缓存变量，作为每一批次处理的变量存储数据#points_single取出对应批次的点云的坐标信息points_single = points[bs_mask][:, 1:4]point_cls_labels_single = point_cls_labels.new_zeros(bs_mask.sum())#将每一个点云数据分配到真实标注框上box_idxs_of_pts = roiaware_pool3d_utils.points_in_boxes_gpu(         points_single.unsqueeze(dim=0), gt_boxes[k:k + 1, :, 0:7].contiguous()).long().squeeze(dim=0)#box_idxs_of_pts是每个点对应分配的标注框索引值，没有匹配的赋值为-1box_fg_flag = (box_idxs_of_pts >= 0) #根据之前扩展的3D框计算被忽略的点if set_ignore_flag:#将每一个点云数据分配到扩展后的标注框上extend_box_idxs_of_pts = roiaware_pool3d_utils.points_in_boxes_gpu(points_single.unsqueeze(dim=0), extend_gt_boxes[k:k+1, :, 0:7].contiguous()).long().squeeze(dim=0)fg_flag = box_fg_flag#异或运算，未扩展前没有包括，扩展后包含到的框，即被忽略的框ignore_flag = fg_flag ^ (extend_box_idxs_of_pts >= 0)point_cls_labels_single[ignore_flag] = -1elif use_ball_constraint:box_centers = gt_boxes[k][box_idxs_of_pts][:, 0:3].clone()box_centers[:, 2] += gt_boxes[k][box_idxs_of_pts][:, 5] / 2ball_flag = ((box_centers - points_single).norm(dim=1) < central_radius)fg_flag = box_fg_flag & ball_flagelse:raise NotImplementedError#记录前景点信息，可以理解为论文中所说的前景点分割gt_box_of_fg_points = gt_boxes[k][box_idxs_of_pts[fg_flag]]#最后一维代表的是标注框对应的类别信息，对应前景点的类别信息point_cls_labels_single[fg_flag] = 1 if self.num_class == 1 else gt_box_of_fg_points[:, -1].long()#记录一次批处理流程中所有点的类别信息point_cls_labels[bs_mask] = point_cls_labels_singleif ret_box_labels and gt_box_of_fg_points.shape[0] > 0:point_box_labels_single = point_box_labels.new_zeros((bs_mask.sum(), 8))#记录每一个前景点从G_hat到G的参数偏移，每个前景点最后输出是1*8维向量fg_point_box_labels = self.box_coder.encode_torch(gt_boxes=gt_box_of_fg_points[:, :-1], points=points_single[fg_flag],gt_classes=gt_box_of_fg_points[:, -1].long())point_box_labels_single[fg_flag] = fg_point_box_labelspoint_box_labels[bs_mask] = point_box_labels_singleif ret_part_labels:point_part_labels_single = point_part_labels.new_zeros((bs_mask.sum(), 3))transformed_points = points_single[fg_flag] - gt_box_of_fg_points[:, 0:3]transformed_points = common_utils.rotate_points_along_z(transformed_points.view(-1, 1, 3), -gt_box_of_fg_points[:, 6]).view(-1, 3)offset = torch.tensor([0.5, 0.5, 0.5]).view(1, 3).type_as(transformed_points)point_part_labels_single[fg_flag] = (transformed_points / gt_box_of_fg_points[:, 3:6]) + offsetpoint_part_labels[bs_mask] = point_part_labels_singletargets_dict = {'point_cls_labels': point_cls_labels,'point_box_labels': point_box_labels,'point_part_labels': point_part_labels}return targets_dict

经典框架解读 | 论文+代码 | 3D Detection | OpenPCDet | PointRCNN