Dynamic InterpIoU: Adaptive IoU Loss

• The paper introduces Dynamic InterpIoU, which extends InterpIoU by adaptively adjusting the interpolation coefficient based on the current IoU to overcome vanishing gradients in non-overlapping cases.
• It replaces handcrafted geometric penalties with an additional IoU term computed on an interpolated box, thereby avoiding undesired enlargement and aligning optimization with precise localization.
• Empirical results on COCO, VisDrone, and PASCAL VOC demonstrate consistent gains in mAP and small object detection, validating the method’s effectiveness across diverse datasets.

Searching arXiv for the target paper and core related IoU-loss papers to ground citations. Searching for "InterpIoU: Rethinking Bounding Box Regression with Interpolation-Based IoU Optimization". Searching for "Generalized Intersection over Union Rezatofighi". Dynamic InterpIoU is a bounding box regression loss for object detection introduced as an extension of InterpIoU in "InterpIoU: Rethinking Bounding Box Regression with Interpolation-Based IoU Optimization" (Liu et al., 16 Jul 2025). It replaces handcrafted geometric penalties with an additional IoU term computed on an interpolated box between the prediction and the ground truth, and it further makes the interpolation coefficient adaptive by tying it to the current IoU value. The method is designed to provide meaningful gradients in non-overlapping cases, avoid the bounding-box enlargement behavior associated with misaligned penalty terms, and improve robustness across object scales and distributions, with experiments on COCO, VisDrone, and PASCAL VOC reporting consistent gains over existing IoU-based losses, particularly for small object detection (Liu et al., 16 Jul 2025).

1. Problem setting and motivation

Bounding box regression is fundamental to object detection, and the regression loss is a primary determinant of localization accuracy. In the formulation underlying Dynamic InterpIoU, the central limitation of standard IoU-based regression is that IoU is non-differentiable in non-overlapping cases in the sense that the gradient of the usual IoU loss vanishes when the predicted box and the ground-truth box do not intersect. Existing IoU-based losses often address this by adding handcrafted geometric penalties.

The motivating critique is that these penalties are sensitive to box shape, size, and distribution. The paper further states that such penalties often lead to suboptimal optimization for small objects and can induce undesired behaviors such as bounding box enlargement because the added penalty is misaligned with the IoU objective. InterpIoU is proposed to remove the handcrafted penalty and instead optimize an additional IoU term on an interpolated box; Dynamic InterpIoU extends this by dynamically adjusting the interpolation coefficient according to the current IoU.

A key implication of this design is that the auxiliary signal is still an IoU objective rather than an externally designed geometric surrogate. The paper’s simulation results are summarized as showing that IoU itself serves as an ideal regression target, while existing geometric penalties are both unnecessary and suboptimal. This suggests a reorientation of bounding box regression away from penalty engineering and toward modifying the optimization path of the IoU target itself.

2. Formal definition of Dynamic InterpIoU

Let the predicted box be

$B_p = [x_p, y_p, w_p, h_p]$

and the ground-truth box be

$B_{gt} = [x_g, y_g, w_g, h_g].$

For a fixed interpolation coefficient $\alpha \in (0,1)$, the interpolated box is defined by a coordinate-wise convex combination:

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$

that is,

$$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$

The standard IoU-based regression loss is

$$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$

InterpIoU augments the usual prediction loss with a second IoU loss on the interpolated box:

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$

or equivalently,

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$

Dynamic InterpIoU makes $\alpha$ data-dependent. Define

$IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$

with clamp bounds $B_{gt} = [x_g, y_g, w_g, h_g].$0 and $B_{gt} = [x_g, y_g, w_g, h_g].$1. The dynamic coefficient is

$B_{gt} = [x_g, y_g, w_g, h_g].$2

The intended operating regime is explicit. When $B_{gt} = [x_g, y_g, w_g, h_g].$3, $B_{gt} = [x_g, y_g, w_g, h_g].$4, which keeps $B_{gt} = [x_g, y_g, w_g, h_g].$5 near $B_{gt} = [x_g, y_g, w_g, h_g].$6 and yields strong early gradients. As $B_{gt} = [x_g, y_g, w_g, h_g].$7, $B_{gt} = [x_g, y_g, w_g, h_g].$8, moving $B_{gt} = [x_g, y_g, w_g, h_g].$9 toward $\alpha \in (0,1)$0 and emphasizing fine alignment. In the low-IoU regime, a large $\alpha \in (0,1)$1 produces a large interpolated overlap and a gradient boost that helps escape flat regions. In the mid- and high-IoU regime, a small $\alpha \in (0,1)$2 makes

$\alpha \in (0,1)$3

thereby focusing optimization on precise localization.

3. Gradient mechanics and optimization behavior

Dynamic InterpIoU is motivated principally by gradient behavior in non-overlapping cases (Liu et al., 16 Jul 2025). When

$\alpha \in (0,1)$4

the standard IoU loss satisfies

$\alpha \in (0,1)$5

This is the vanishing-gradient failure mode that motivates the method.

If $\alpha \in (0,1)$6 is sufficiently large that the interpolated box overlaps the target, then

$\alpha \in (0,1)$7

and by the chain rule,

$\alpha \in (0,1)$8

Thus $\alpha \in (0,1)$9 controls the strength of the rescue gradient. Larger $B_{int} = (1-\alpha) B_p + \alpha B_{gt}$0 improves the likelihood that $B_{int} = (1-\alpha) B_p + \alpha B_{gt}$1 overlaps $B_{int} = (1-\alpha) B_p + \alpha B_{gt}$2, but it also reduces the backpropagated scale through the factor $B_{int} = (1-\alpha) B_p + \alpha B_{gt}$3. Dynamic InterpIoU is designed to manage this trade-off automatically rather than with a single static interpolation coefficient.

The paper distinguishes overlapping and non-overlapping regimes. In the overlapping case, both

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$4

contribute, guiding both center and size. In the non-overlapping case,

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$5

but

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$6

through the interpolated term and the Jacobian

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$7

The paper also analyzes why the method avoids box enlargement. For width $B_{int} = (1-\alpha) B_p + \alpha B_{gt}$8, the standard IoU gradient is given as

$B_{int} = (1-\alpha) B_p + \alpha B_{gt}$9

with $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$0, $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$1, and $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$2. The second term always penalizes area growth because $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$3 and $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$4, so any change in $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$5 that does not increase the intersection is driven negative. Because InterpIoU is the sum of two IoU losses, its net gradient inherits this no-enlargement property. A plausible implication is that Dynamic InterpIoU preserves the same mechanism while modulating how strongly the auxiliary IoU term contributes across training stages.

4. Relation to handcrafted geometric penalties

Many earlier IoU-based losses add a penalty term $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$6 based on shape or distance rather than on IoU alone. The paper lists several representative examples.

Loss	Penalty term	Notes stated in the paper
GIoU	$$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$7	Minimizing it can enlarge $$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$8 to fill the enclosing box
DIoU	$$B_{int} = [ (1-\alpha)x_p+\alpha x_g,\; (1-\alpha)y_p+\alpha y_g,\; (1-\alpha)w_p+\alpha w_g,\; (1-\alpha)h_p+\alpha h_g ].$$9	Uses center distance and enclosing-box diagonal
CIoU	$$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$0	Adds aspect-ratio term $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$1 with weight $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$2
EIoU, SIoU, PIoU	Not expanded in the provided details	Listed as related IoU-based losses

For GIoU, the enclosing quantity $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$3 is the area of the smallest enclosing box. The paper states that minimizing $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$4 can be achieved by enlarging the predicted box to fill the enclosing box, and it identifies this as a source of box enlargement. In controlled simulations, these penalties often produce

$$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$5

even when overlap is zero, thereby driving the predicted box to grow.

To quantify enlargement, the paper defines

$$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$6

According to the reported simulations, for GIoU, CIoU, DIoU, and SIoU this metric regularly exceeds $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$7, interpreted in the paper as 600% enlargement, whereas InterpIoU keeps $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$8. The central methodological distinction is therefore not merely that Dynamic InterpIoU adds another term, but that its additional term remains an IoU objective evaluated at an interpolated box instead of a handcrafted geometric correction.

A common misconception in this area is that non-overlap necessarily requires an explicit geometric penalty. The paper’s position is narrower and more specific: the problem is addressed by constructing an interpolated box that restores overlap and therefore restores an IoU gradient, without introducing a separate penalty objective. This does not eliminate geometric considerations from optimization, but relocates them into the interpolation path between $$\mathcal{L}_{IoU}(B_a,B_b)=1-\mathrm{IoU}(B_a,B_b).$$9 and $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$0.

5. Training procedure and operating regimes

Dynamic InterpIoU is implemented with a short training loop. For each training iteration, the detector predicts $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$1 for each matched anchor or box, computes

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$2

sets

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$3

constructs

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$4

and forms

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$5

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$6

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$7

Backpropagation then updates the network weights using $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$8.

The paper reports that in most experiments the static version, InterpIoU, uses $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = \mathcal{L}_{IoU}(B_p,B_{gt}) + \mathcal{L}_{IoU}(B_{int},B_{gt}),$$9. For Dynamic InterpIoU, the clamp ranges are dataset-specific: $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$0 for VOC, $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$1 for VisDrone, and $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$2 for COCO. Training settings are also specified: SSD uses 140 k iterations with batch 32; YOLOv8 uses 150 epochs on VOC and VisDrone and 500 epochs on COCO with batch 16/4; DINO uses the official 12 epochs with batch 8.

The reported practical recommendations are correspondingly direct. If a dataset has consistent object scales, a fixed $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$3 works robustly. For mixed-scale regimes or to boost early-stage learning, the recommended variant is D-InterpIoU with clamp approximately $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$4. The loss can replace an existing IoU-based loss term directly, and no additional hyperparameters beyond $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$5 and $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$6 are needed. The paper also notes an implementation constraint arising from the factor

$$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$7

namely that very large $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$8 yields a small gradient on $$\mathcal{L}_{InterpIoU}(B_p,B_{gt};\alpha) = [1-\mathrm{IoU}(B_p,B_{gt})] + [1-\mathrm{IoU}(B_{int},B_{gt})].$$9, so $\alpha$0 should be avoided.

6. Empirical results and ablation behavior

The empirical evaluation covers PASCAL VOC 2007, VisDrone, and MS COCO across SSD, YOLOv8-m, and DINO-4scale (Liu et al., 16 Jul 2025). The benchmarks are selected to span medium-sized objects, small dense aerial objects, and a wide scale distribution, respectively.

On PASCAL VOC with SSD and YOLOv8, InterpIoU improves mAP by $\alpha$1 to $\alpha$2 over the best prior IoU loss, identified in the paper as SIoU or PIoU. D-InterpIoU adds a further $\alpha$3 on small classes such as bottle and pottedplant.

On VisDrone with YOLOv8, the baseline $\alpha$4 is reported as $\alpha$5, InterpIoU as $\alpha$6, and D-InterpIoU as $\alpha$7. The paper also reports $\alpha$8 gains of $\alpha$9 and an overall mAP gain of $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$0.

On MS COCO with YOLOv8-m, the best baseline mAP is $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$1 using SIoU, with $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$2. InterpIoU reaches $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$3 with $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$4, while D-InterpIoU improves small-object performance to $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$5, a gain of $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$6. On DINO-4scale, the best baseline mAP is $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$7 using PIoU, and InterpIoU reaches $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$8 with $IoU_0 = \mathrm{IoU}(B_p,B_{gt}),$9, also reported as a $B_{gt} = [x_g, y_g, w_g, h_g].$00 gain on small objects.

The ablation on $B_{gt} = [x_g, y_g, w_g, h_g].$01 further clarifies the operating regime. On VisDrone and VOC with YOLOv8-s, static $B_{gt} = [x_g, y_g, w_g, h_g].$02 in the range $B_{gt} = [x_g, y_g, w_g, h_g].$03 yields a steady mAP increase as $B_{gt} = [x_g, y_g, w_g, h_g].$04, while performance plateaus for $B_{gt} = [x_g, y_g, w_g, h_g].$05. The best static setting is $B_{gt} = [x_g, y_g, w_g, h_g].$06, with VisDrone mAP $B_{gt} = [x_g, y_g, w_g, h_g].$07 and VOC mAP $B_{gt} = [x_g, y_g, w_g, h_g].$08. A dynamic setting with clamp $B_{gt} = [x_g, y_g, w_g, h_g].$09 yields VisDrone mAP $B_{gt} = [x_g, y_g, w_g, h_g].$10 and VOC mAP $B_{gt} = [x_g, y_g, w_g, h_g].$11, together with $B_{gt} = [x_g, y_g, w_g, h_g].$12 gains. These results are consistent with the method’s stated rationale: large interpolation is beneficial for escaping poor initial localization, whereas reduced interpolation is preferable for fine-grained final alignment.