corner

Corner loss of the box-refinement heads (PointRCNN).

Classes:

CornerLoss –

Module form of corner_loss.

CornerLoss [source] ¶

CornerLoss(
    beta: float = 1.0,
    reduction: Literal["none", "sum", "mean"] = "mean",
)

Bases: Module

Module form of corner_loss.

Parameters:

beta (float, default: 1.0 ) –

Smooth-\(L_1\) transition point.
reduction (Literal['none', 'sum', 'mean'], default: 'mean' ) –

"none" returns the per-box loss, "sum" its sum, "mean" its mean.

Example

import torch
from torch_pointcloud.losses import CornerLoss

criterion = CornerLoss()
box = torch.tensor([[0.0, 0.0, 0.0, 4.0, 2.0, 1.5, 0.3]])
criterion(box, box).item()  # 0.0

Methods:

forward –

Compute the corner loss between matched boxes pred_boxes \((N, 7)\) and gt_boxes \((N, 7)\).

forward [source] ¶

forward(pred_boxes: Tensor, gt_boxes: Tensor) -> Tensor

Compute the corner loss between matched boxes pred_boxes \((N, 7)\) and gt_boxes \((N, 7)\).

corner_loss [source] ¶

corner_loss(
    pred_boxes: Tensor,
    gt_boxes: Tensor,
    *,
    beta: float = 1.0,
) -> Tensor

Per-box mean smooth-\(L_1\) distance between the eight corners of two box sets, robust to the heading flip.

The corner regularization of PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud (Shi et al., 2019): each predicted corner is compared with the matching corner of the ground-truth box and of that box rotated by \(\pi\), and the closer of the two distances is penalized, so a heading off by half a turn is not punished as a position error.

Parameters:

pred_boxes (Tensor) –

Predicted boxes \((c_x, c_y, c_z, d_x, d_y, d_z, \theta)\), shape \((N, 7)\).
gt_boxes (Tensor) –

Matched ground-truth boxes, shape \((N, 7)\).
beta (float, default: 1.0 ) –

Smooth-\(L_1\) transition point.

Returns:

Tensor –

The per-box loss, shape \((N,)\).

Shape

pred_boxes: \((N, 7)\)
gt_boxes: \((N, 7)\)
output: \((N,)\)

Example

>>> box = torch.tensor([[0.0, 0.0, 0.0, 4.0, 2.0, 1.5, 0.3]])
>>> corner_loss(box, box).tolist()
[0.0]
>>> flipped = box.clone()
>>> flipped[:, 6] += math.pi
>>> corner_loss(box, flipped).abs().max().item() < 1e-5
True