DETR Hungarian matching and set-prediction loss (Geron Ch 16, grdetr)

Cost C(i,j) = -p_i(c_j) + lam_bbox ||b_i - b_j||_1 - lam_giou GIoU, minimised exactly by the Jonker-Volgenant shortest-augmenting-path assignment. matching pairs are 0-based (pred, gt), as in Python.

Usage

morie_geron_detr_hungarian_matching(
  pred_boxes,
  pred_classes,
  gt_boxes,
  gt_classes,
  lam_bbox = 5,
  lam_giou = 2,
  box_format = "xyxy",
  no_object_class = NULL,
  eos_coef = 0.1,
  class_is_logits = TRUE
)

Arguments

pred_boxes, gt_boxes

Box matrices (N, 4) / (M, 4).

pred_classes

Logits or probabilities (N, C).

gt_classes

0-based ground-truth classes.

lam_bbox, lam_giou

Non-negative cost weights.

box_format

"xyxy" or "cxcywh".

no_object_class

Optional 0-based no-object index.

eos_coef

No-object down-weight.

class_is_logits

Softmax needed.

Value

List with matching, cost_matrix, total_cost, loss, loss_class, loss_bbox, loss_giou, loss_no_object, matched_giou, matched_l1, unmatched_predictions.

Examples

pb <- matrix(c(0, 0, 1, 1, 10, 10, 11, 11, 0, 0, 2, 2), 3, 4,
    byrow = TRUE)
pc <- matrix(c(10, 0, 0, 10, 1, 1), 3, 2, byrow = TRUE)
gb <- matrix(c(0, 0, 1, 1, 10, 10, 11, 11), 2, 4, byrow = TRUE)
morie_geron_detr_hungarian_matching(pb, pc, gb, c(0, 1), no_object_class = 1)
#> $matching
#> $matching[[1]]
#> [1] 0 0
#> 
#> $matching[[2]]
#> [1] 1 1
#> 
#> 
#> $cost_matrix
#>            [,1]       [,2]
#> [1,]  -2.999955 201.966897
#> [2,] 201.966897  -2.999955
#> [3,]   9.000000 191.417355
#> 
#> $total_cost
#> [1] -5.999909
#> 
#> $loss
#> [1] 0.06940552
#> 
#> $loss_class
#> [1] 9.07978e-05
#> 
#> $loss_bbox
#> [1] 0
#> 
#> $loss_giou
#> [1] 0
#> 
#> $loss_no_object
#> [1] 0.06931472
#> 
#> $matched_giou
#> [1] 1 1
#> 
#> $matched_l1
#> [1] 0 0
#> 
#> $unmatched_predictions
#> [1] 2
#> 
#> $estimate
#> [1] 0.06940552
#> 
#> $n
#> [1] 2
#> 
#> $method
#> [1] "DETR Hungarian matching and set-prediction loss"
#>