FP-IoU Loss: Overview of IoU-Based Overlap Optimization

• FP-IoU Loss Function is a term for a family of IoU-based objectives that maximize overlap between predictions and ground truth rather than minimizing coordinate errors.
• It encompasses diverse formulations such as plain IoU loss, misprediction-set Jaccard loss, and focal or power reweightings, each adapted for different settings like tracking, segmentation, or detection.
• These formulations have practical applications in tracking, segmentation, object detection, and crop mapping, offering tailored solutions for precision, gradient behavior, and imbalance handling.

FP-IoU loss function denotes, in the literature surveyed here, an ambiguous label around Intersection over Union (IoU) or Jaccard-derived objectives that optimize overlap rather than coordinate error. A central point is that several relevant papers explicitly do not define a variant called FP-IoU; instead they use plain IoU between predicted and ground-truth boxes or masks, or introduce adjacent formulations such as misprediction-set Jaccard losses, class-wise soft IoU, focal or power reweightings, interval-mapped IoU, smoothed IoU, and context-conditioned IoU penalties (Dasari et al., 2020). This suggests that “FP-IoU” is better treated as a non-standard umbrella expression than as a single canonical formula, with its meaning depending on whether the setting is tracking, semantic segmentation, 2D/3D object detection, or small-object localization (Mohammadi et al., 2021).

1. Terminological status and scope

In the cited arXiv literature, FP-IoU is not a standardized technical name. The IG-Track paper states that it “does not define any variant called FP-IoU, foreground-foreground IoU, foreground-proposal IoU, or any similarly named decomposition,” and instead uses plain IoU between the final predicted bounding box $pred\_bb$ and the ground-truth bounding box $gt\_bb$ (Dasari et al., 2020). The 3D FCN crop-mapping paper likewise states that it does not use the term FP-IoU and that its loss is best characterized as a standard class-wise IoU or Jaccard loss summed over crop classes (Mohammadi et al., 2021). The Focaler-IoU paper makes the same point in a different way: the only formal name introduced there is Focaler-IoU, not FP-IoU (Zhang et al., 2024).

A different usage appears in the summary of “Alpha-IoU,” where the FP-IoU idea is described as essentially the generalized power-IoU formulation rather than a separate unrelated objective (He et al., 2021). Even there, however, the operative formalism is the $\alpha$-IoU family, not a distinct symbol or universally adopted FP-IoU notation. The most precise encyclopedia-level characterization is therefore negative: FP-IoU is not a single loss with a fixed formula across the literature. It is a label that often points toward IoU-based objectives that make false positives, false negatives, or high/low-overlap regimes more explicit, but the underlying mathematics varies substantially.

This terminological instability matters because multiple non-equivalent constructions can appear similar at a verbal level. Standard IoU regression, soft mask IoU, IoU-balanced weighting, Focal-EIoU, Focaler-IoU, $\alpha$-IoU, and smoothed IoU all optimize overlap-related quantities, yet they differ in whether they modify the overlap term itself, reweight samples, add geometric penalties, reshape gradients, or inject contextual constraints (Wu et al., 2019).

2. Canonical overlap-maximization formulations

The simplest and most recurrent formulation is the direct loss

$L = 1 - IoU,$

or an obvious class-wise extension of it. In IG-Track, the explicit new term is

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$

and the full objective is

$Loss = L_{cls} + L_{reg} + L_{iou}.$

The paper emphasizes that the objective directly maximizes overlap between the final predicted box and ground truth, rather than regressing box coordinates with an $L_1$, $L_2$, or Smooth-$L_1$ distance (Dasari et al., 2020).

In segmentation, the same idea is written with soft masks. The CT organ segmentation paper defines

$gt\_bb$0

where $gt\_bb$1 is the predicted voxel probability vector and $gt\_bb$2 is the binary ground-truth vector (Rister et al., 2018). For multi-class crop mapping, the 3D FCN paper defines a class-wise sum

$gt\_bb$3

with $gt\_bb$4 computed from predicted and reference masks for class $gt\_bb$5 (Mohammadi et al., 2021).

For 2D and 3D object detection, the same canonical loss appears at the box level. “IoU Loss for 2D/3D Object Detection” defines the usual Jaccard overlap and uses

$gt\_bb$6

extending the computation to rotated 2D boxes and 3D boxes through bird’s-eye-view overlap plus height overlap (Zhou et al., 2019).

Setting	Representative loss	Brief characterization
Visual tracking	$gt\_bb$7	Final predicted box vs. ground truth (Dasari et al., 2020)
CT organ segmentation	$gt\_bb$8	Soft Jaccard on voxel probabilities (Rister et al., 2018)
Crop mapping	$gt\_bb$9	Class-wise IoU summed over classes (Mohammadi et al., 2021)
2D/3D detection	$\alpha$0	Bounding-box overlap loss (Zhou et al., 2019)

Across these settings, the shared rationale is that overlap quality is the actual downstream criterion. Tracking papers contrast this with inference-time heuristics and coordinate residual losses; segmentation papers contrast it with per-pixel cross-entropy; object detection papers contrast it with $\alpha$1- or $\alpha$2-style parameter regression (Dasari et al., 2020).

3. Mispredictions, false positives, and set-function theory

The most rigorous theoretical clarification appears in “Yes, IoU loss is submodular - as a function of the mispredictions.” That paper fixes a finite base set $\alpha$3, a ground-truth subset $\alpha$4, and a prediction subset $\alpha$5, and then distinguishes two different set functions. As a function of the prediction set,

$\alpha$6

the Jaccard index is, in general, neither submodular nor supermodular. The key reparameterization introduces the symmetric difference

$\alpha$7

which is the set of mispredictions, and rewrites the same quantity as

$\alpha$8

The paper’s central claim is that $\alpha$9 is supermodular, and therefore $\alpha$0 is submodular (Berman et al., 2018).

This distinction is directly relevant to any interpretation of FP-IoU that emphasizes false positives and false negatives. The paper’s point is not that IoU is submodular in the predicted set itself, but that IoU/Jaccard has the required set-function structure when expressed in terms of the error set. It also argues that a criticism in prior work arose from conflating the prediction-set function with the misprediction-set function (Berman et al., 2018).

A related, but more application-oriented, analysis appears in the CT organ segmentation work. There the authors derive approximate penalties for false negatives and false positives under IoU loss: $\alpha$1 where $\alpha$2 and $\alpha$3. When $\alpha$4, both are approximately equal. The paper uses this to argue that IoU loss produces approximately balanced penalties for FP and FN when predictions are already reasonably close to the target, unlike weighted cross-entropy, whose per-voxel weighting depends on class size (Rister et al., 2018).

Taken together, these papers do not define a named FP-IoU loss, but they do supply the two main theoretical lenses from which such a phrase is usually interpreted: first, IoU as a function of the misprediction set rather than the prediction set; second, IoU as a loss whose denominator implicitly aggregates true positives, false positives, and false negatives through the union term (Mohammadi et al., 2021).

4. Segmentation and crop-mapping formulations

In segmentation-oriented uses, IoU losses are typically class-wise or mask-wise and are motivated by the mismatch between overlap-based evaluation and voxelwise or pixelwise training losses. The crop-mapping paper explicitly argues that cross-entropy optimizes per-pixel classification independently, whereas the task is inherently a segmentation or areal mapping problem. Its loss therefore directly increases the overlap between the prediction map and ground reference mask: $\alpha$5 The denominator term $\alpha$6 is described as the union term, so the formula implicitly captures true positives via $\alpha$7, false positives via $\alpha$8 when $\alpha$9, and false negatives via $L = 1 - IoU,$0 when $L = 1 - IoU,$1, even though the paper does not introduce FP-IoU terminology (Mohammadi et al., 2021).

The same paper reports that switching from 3DFCN + CE loss to 3DFCN + IOU loss changes Kappa from 91.3 to 91.8, MA-PA from 93.7 to 94.1, and MA-UA from 93.6 to 94.2. The conclusion drawn there is that IoU loss is a better choice for learning individual crop types because it directly optimizes overlap (Mohammadi et al., 2021).

In medical CT segmentation, the argument is both mathematical and empirical. The IOU loss is presented as a close relative of the Dice loss, but the paper emphasizes that IoU/Jaccard is a metric on binary sets whereas Dice is not. It also argues that there is no unique IOU loss function, and introduces a family

$L = 1 - IoU,$2

as well as a more general family with monotone transforms $L = 1 - IoU,$3 (Rister et al., 2018).

Empirically, that paper reports that IOU loss outperforms cross-entropy for all listed organs, both with and without augmentation. For example, bladder Dice changes from 70.4 with CE and no augmentation to 80.1 with IOU and no augmentation, and to 83.7 with IOU plus augmentation. On the LiTS liver benchmark, average Dice changes from 0.866 with cross-entropy and no augmentation to 0.905 with IOU and augmentation (Rister et al., 2018). These results are not presented as evidence for a separate FP-IoU formula; rather, they show that standard soft IoU formulations already have favorable imbalance properties in segmentation.

5. Bounding-box regression variants in detection and tracking

Bounding-box regression has produced the broadest family of IoU-derived objectives, and most of the formulations that might be informally called FP-IoU belong here. The direct baseline remains $L = 1 - IoU,$4, but the literature modifies it in several distinct ways.

One direction reweights standard losses using IoU. “IoU-balanced Loss Functions for Single-stage Object Detection” defines an IoU-balanced classification loss and an IoU-balanced localization loss. The localization branch uses

$L = 1 - IoU,$5

with

$L = 1 - IoU,$6

The paper’s interpretation is that high-IoU positives contribute more to regression learning and low-IoU positives contribute less, thereby suppressing gradient domination by poorly localized examples. On COCO test-dev, the reported improvement is AP by $L = 1 - IoU,$7 and AP75 by $L = 1 - IoU,$8 (Wu et al., 2019).

A second direction changes the geometry term itself. “Focal and Efficient IOU Loss for Accurate Bounding Box Regression” introduces EIOU,

$L = 1 - IoU,$9

and then applies IoU-based focal weighting,

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$0

The stated purpose is to focus regression on high-quality anchor boxes and suppress low-quality ones (Zhang et al., 2021).

A third direction reshapes the IoU scale itself. Focaler-IoU defines a piecewise linear remapping

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$1

and then

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$2

Its central claim is that different tasks may be dominated by easy or difficult regression samples, and that the interval $IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$3 can be adjusted to emphasize the desired regime. On VOC with YOLOv8 plus SIoU, the paper reports +0.3 on both AP50 and mAP50:95, while on the tiny-object AI-TOD dataset with YOLOv5 plus SIoU it reports +1.9 AP50 and +0.5 mAP50:95 (Zhang et al., 2024).

A fourth direction is the power-family view. “Alpha-IoU” generalizes IoU-based regression to

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$4

and extends the same power transformation to GIoU, DIoU, and CIoU. The paper emphasizes order preservingness and loss or gradient reweighting, and reports that $IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$5 works well overall, with $IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$6 as the most robust default in its experiments (He et al., 2021).

A fifth direction addresses the non-smooth or flat landscape away from overlap. “Intersection over Union with smoothing for bounding box regression” proposes

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$7

where the smoothing term is constructed from left and right offsets and position-dependent weights so that supervision increases from the ground-truth box toward the image border. On the reported dataset, test-average IoU similarity changes from 0.536 for IoU to 0.684 for IoU + smooth, with a best test result of 0.729 (Števuliáková et al., 2023).

Finally, recent small-object work adds inter-class context rather than only box geometry. The inter-class relational loss defines

$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$8

with the predicted small-object box expected to lie inside a related large-object box. The paper presents this as a plug-and-play wrapper around IoU, GIoU, DIoU, or CIoU, and reports +10.3% mAP$IOU\_loss = 1- IOU(pred\_bb, gt\_bb),$9 and +1.6% mAP$Loss = L_{cls} + L_{reg} + L_{iou}.$0 for YOLOv12-T on its test set (Ning et al., 20 Aug 2025).

6. Misconceptions, limitations, and current interpretation

A common misconception is that FP-IoU names a single foreground/proposal decomposition. The papers surveyed here repeatedly contradict that reading. IG-Track explicitly states that it does not introduce separate foreground/proposal terms and that the only explicit optimization target is

$Loss = L_{cls} + L_{reg} + L_{iou}.$1

The crop-mapping paper likewise states that there is no separate FP-IoU term and no explicit decomposition into true positives, false positives, and false negatives in the text, even though the union denominator has exactly that structure (Dasari et al., 2020).

A second misconception is that all theoretical claims about IoU submodularity concern the prediction set directly. The set-function note shows that this is incorrect: the Jaccard index is, in general, neither submodular nor supermodular as a function of the prediction set $Loss = L_{cls} + L_{reg} + L_{iou}.$2, and the relevant property appears only after reparameterization by the misprediction set $Loss = L_{cls} + L_{reg} + L_{iou}.$3 (Berman et al., 2018).

A third misconception is that plain IoU is always sufficient once overlap is optimized directly. Several papers identify concrete limitations: no meaningful gradient when boxes do not overlap, flat gradients for small objects, or poor alignment between classification scores and localization quality. The remedies differ. Some papers add geometric penalties, as in EIOU or GIoU; some reweight high- and low-IoU examples, as in IoU-balanced loss, Focal-EIOU, Focaler-IoU, and $Loss = L_{cls} + L_{reg} + L_{iou}.$4-IoU; some smooth the loss surface over the spatial search space; and some use inter-class context to guide small-object predictions (Zhou et al., 2019).

A further limitation is evidential rather than mathematical. IG-Track reports improvements on VOT2018 and GOT-10k, including an EAO change from 0.290 to 0.327 and SR$Loss = L_{cls} + L_{reg} + L_{iou}.$5 from 0.195 to 0.220, but it also states that it does not provide an ablation table isolating the loss alone from the IOU module or from other design choices (Dasari et al., 2020). This cautions against equating a reported performance gain with a uniquely identifiable FP-IoU mechanism.

The current state of the literature therefore supports a restrained conclusion. FP-IoU is not a settled standalone loss name. What the literature does provide is a coherent family of IoU-centered constructions: direct overlap losses for boxes and masks, misprediction-set formulations with submodular structure, class-wise soft IoU objectives, IoU-based weighting schemes, power-law and focal remappings, smoothed variants, and contextual multiplicative penalties. Any precise use of the term requires the accompanying formula, because the surrounding literature does not supply a universal one (He et al., 2021).