SeaBird: BEV Segmentation for 3D Large Objects

• SeaBird is a monocular 3D object detection method that employs foreground BEV segmentation supervised with Dice loss to improve localization of large objects.
• It reformulates depth regression into a BEV segmentation problem, mitigating noise sensitivity and addressing traditional losses that struggle with large-object scales.
• Empirical evaluations on KITTI-360 and nuScenes demonstrate significant AP improvements on large objects, validating the sequential BEV segmentation and detection strategy.

SeaBird is a monocular 3D object-detection method in which bird’s-eye-view (BEV) foreground segmentation, supervised with Dice loss, is inserted as an intermediate representation for downstream 3D box prediction. Introduced in “SeaBird: Segmentation in Bird’s View with Dice Loss Improves Monocular 3D Detection of Large Objects” and later expanded in a broader thesis on monocular 3D generalization, SeaBird targets a specific failure mode of Mono3D systems: strong performance on cars and smaller objects, but severe degradation on large objects such as trucks, buses, and trailers. Its central claim is that this degradation is not explained only by class imbalance or receptive-field limitations; rather, it is substantially tied to the noise sensitivity of conventional depth-regression losses. SeaBird addresses that failure mode by converting localization into a BEV segmentation problem whose Dice-based supervision becomes increasingly favorable as object extent grows (Kumar et al., 2024, Kumar, 27 Aug 2025).

1. Origin, scope, and problem domain

SeaBird belongs to the literature on monocular 3D object detection from images, especially in autonomous-driving settings where the task is to estimate an object’s class, 3D position, dimensions, and orientation from camera input. In the 2024 paper, SeaBird is presented as a method for improving large-object detection in camera-only pipelines; in the 2025 thesis, it is positioned explicitly as the “large-object generalization” contribution alongside other components addressing occlusion, dataset shift, and camera-height robustness (Kumar et al., 2024, Kumar, 27 Aug 2025).

The method is not introduced as a wholly new image-to-BEV backbone. Instead, it is a plugin-style BEV augmentation that operates on systems already capable of producing BEV features from images. On KITTI-360, it is instantiated as Image2Maps + SeaBird and PanopticBEV + SeaBird; on nuScenes, it is integrated into BEVerse and HoP. The paper therefore treats SeaBird less as a standalone detector than as a design principle: use a BEV foreground occupancy representation, trained with Dice loss, to stabilize the localization stage that frontal Mono3D methods often treat as direct metric regression (Kumar et al., 2024).

A common source of confusion is the name itself. In the broader arXiv literature, “seabird” frequently refers to ecological monitoring or bird reconstruction problems, including monocular 3D reconstruction of common murres from video. SeaBird, by contrast, denotes a BEV-segmentation-guided Mono3D method and is unrelated to seabird ethology or wildlife monitoring as such (Hägerlind et al., 2024).

2. Failure diagnosis: why large objects are difficult in Mono3D

SeaBird begins from an empirical observation: modern frontal-view monocular detectors perform reasonably on cars and smaller categories, yet often fail badly on large objects. The 2024 paper argues that this is an understudied generalization problem, while the later thesis sharpens the diagnosis by noting that poor large-object performance persists even on KITTI-360, which is described as a nearly balanced large-object benchmark, with large objects and cars present in roughly a $1:2$ ratio (Kumar et al., 2024, Kumar, 27 Aug 2025).

This empirical setup is important because it weakens two common explanations. The first is data scarcity or class imbalance. The second is insufficient receptive field. SeaBird does not deny that either factor may matter, but it argues they are incomplete. Even when large objects are not extremely rare, frontal detectors still exhibit near-zero large-object AP on KITTI-360 validation. The method therefore reframes the problem as one of loss geometry under depth noise.

The paper’s simplified model assumes additive depth error,

$$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$

and interprets the resulting 3D failure through BEV displacement along the viewing ray. In frontal Mono3D pipelines, large objects are usually supervised by regression losses such as $L_1$, $L_2$, or Smooth-$L_1$ on depth or related coordinates. SeaBird’s diagnosis is that these losses remain highly sensitive to additive depth noise even when the underlying object is large, whereas overlap-based BEV supervision becomes relatively more tolerant because a fixed metric displacement destroys a much smaller fraction of the footprint of a long object than of a short one (Kumar et al., 2024).

The thesis makes this claim more explicit by arguing that large-object failure is partly an optimization-and-loss issue. In that formulation, the critical variable is not merely whether a detector can “see” a trailer or bus, but whether the supervision applied to localization is robust enough to converge toward stable solutions under realistic depth uncertainty (Kumar, 27 Aug 2025).

3. Representation, architecture, and training protocol

SeaBird’s architectural move is to place a foreground BEV segmentation head in front of the 3D detector. Image features are first transformed into latent BEV features. A segmentation head then predicts a BEV map of foreground detection categories only. That predicted segmentation map is concatenated with the original BEV features and passed to a Mono3D detection head, which outputs a 7-DoF box parameterization consisting of BEV 2D position, elevation, 3D dimensions, and yaw (Kumar et al., 2024, Kumar, 27 Aug 2025).

The design is explicitly sequential rather than parallel. The authors argue that naive parallel multitask learning with BEV segmentation and detection can hurt detection through negative transfer. SeaBird’s mitigation has two parts. First, the segmentation target is restricted to foreground detection categories, not all semantic classes. Second, the system is trained in two stages so that the segmentation head can first learn a Dice-supervised localization prior before the detector is jointly optimized to consume it (Kumar, 27 Aug 2025).

On KITTI-360, the SeaBird extensions are built on Image2Maps and PanopticBEV, with BoxNet used as the detector head. On nuScenes, BEVerse is reorganized into the sequential pattern, while HoP imports a BEV segmentation head from BEVerse. Across these variants, SeaBird does not replace box prediction with segmentation alone. The segmentation branch contributes BEV localization structure, but the detection branch still predicts additional 3D attributes, especially height and elevation, which are not recoverable from the segmentation map by itself (Kumar, 27 Aug 2025).

The training protocol is likewise central. Stage one optimizes the segmentation head with Dice loss. Stage two jointly fine-tunes segmentation and detection: $L = L_{det} + \lambda_{seg} L_{seg}.$ For most settings, the paper uses $\lambda_{seg}=5$; where an implementation internally scales segmentation loss, the effective coefficient is adjusted accordingly, such as $\lambda_{seg}=35$ in one PanopticBEV setting. The KITTI-360 implementations supervise BEV segmentation with Dice loss, BEV-space localization and size terms with Smooth-$L_1$, and orientation with cross-entropy. The paper also shows that a BEV feature shortcut is necessary: passing only the predicted segmentation map to the detector reduces performance because segmentation alone lacks the vertical information required for elevation estimation (Kumar et al., 2024).

4. Dice loss as a large-object localization objective

SeaBird’s most distinctive contribution is its explicit comparison between regression losses and Dice loss under a simplified noise model. The paper adopts an SGD convergence lens in which

$$E\left(\|\hat{\mathbf w}-\mathbf w_*\|_2^2\right) = c_1 \, \mathrm{Var}(g_\epsilon) + c_2,$$

where $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$0. Under this analysis, lower gradient variance implies a converged solution closer to the optimum (Kumar et al., 2024).

For the standard regression baselines, the paper gives: $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$1 and

$$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$2

Neither variance term improves merely because the underlying object is longer (Kumar et al., 2024).

SeaBird then analyzes a one-class BEV localization setup in which a predicted footprint of length $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$3 is shifted by $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$4 along the ray. The Dice loss becomes

$$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$5

with gradient

$$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$6

Its gradient variance is

$$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$7

The critical property is the $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$8 dependence: for fixed noise magnitude, Dice becomes more favorable as object length increases (Kumar et al., 2024).

The thesis reformulates the same argument in terms of large-object robustness and gives concrete examples. For cars with $$\hat d = d + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$9 m, the reported threshold beyond which Dice has better convergence is about 0.3 m of depth-noise deviation; for trailers with $L_1$0 m, it drops to about 0.1 m. The intended interpretation is that the larger the object, the earlier Dice becomes the preferable supervision signal under depth uncertainty (Kumar, 27 Aug 2025).

This theoretical analysis is explicitly limited. It assumes a linear model, additive Gaussian depth noise, perfect foreground classification in the one-class setup, and even a theorem in which depth is the only source of error. The authors present it as an explanatory model rather than a full theory of modern end-to-end Mono3D systems (Kumar, 27 Aug 2025).

5. Empirical performance on KITTI-360 and nuScenes

SeaBird is evaluated primarily on KITTI-360 and nuScenes, with the most visible gains concentrated in the large-object regime. On KITTI-360 validation under the stricter metric, frontal monocular baselines such as GrooMeD-NMS, MonoDLE, GUP Net, DEVIANT, Cube R-CNN, and MonoDETR achieve large-object AP values near zero, ranging from 0.00 to 0.94. SeaBird shifts this regime substantially: I2M+SeaBird reaches Lrg 8.71, while PBEV+SeaBird reaches Lrg 13.22. Under the looser metric, the same variants reach 35.76 and 37.15 on the large category, respectively (Kumar et al., 2024, Kumar, 27 Aug 2025).

On the official KITTI-360 test benchmark, the thesis reports that frontal monocular baselines remain below 1 on one reported metric and below 29 on the other, whereas I2M+SeaBird reaches 3.14 and 35.04, and PBEV+SeaBird reaches 4.64 and 37.12. The authors describe these results as state of the art for monocular camera-only methods on that benchmark, and note that PBEV+SeaBird surpasses older LiDAR baselines on the metrics shown (Kumar, 27 Aug 2025).

On nuScenes, SeaBird acts as a consistent improvement rather than a one-dataset artifact. Gains are reported across BEVerse and HoP variants, usually with the largest improvements on the Lrg group. The strongest example in the supplied results is HoP V2-99, where adding SeaBird improves Lrg 36.5 \rightarrow 40.3, mAP 49.6 \rightarrow 52.7, and NDS 58.3 \rightarrow 60.2 on validation. Test-set improvements are also reported, including HoP V2-99 moving from Lrg 37.1, mAP 49.4, NDS 58.9 to Lrg 38.4, mAP 51.1, NDS 59.7 (Kumar et al., 2024, Kumar, 27 Aug 2025).

Setting	Baseline regime	SeaBird result
KITTI-360 val, strict metric	Frontal Mono3D large-object AP mostly $L_1$1–$L_1$2	I2M+SeaBird $L_1$3, PBEV+SeaBird $L_1$4
KITTI-360 test	Monocular baselines below $L_1$5 on one metric and below $L_1$6 on the other	I2M+SeaBird $L_1$7, PBEV+SeaBird $L_1$8
nuScenes val, HoP V2-99	Lrg $L_1$9, mAP $L_2$0, NDS $L_2$1	Lrg $L_2$2, mAP $L_2$3, NDS $L_2$4

A particularly important validation is the lengthwise analysis on KITTI-360. The paper states that SeaBird pipelines outperform all baselines on objects over $L_2$5 m in length, directly matching the size-dependent prediction of the Dice analysis (Kumar et al., 2024).

6. Ablations, interpretation, and limitations

The ablation studies show that SeaBird’s performance is not attributable to BEV representation alone. On KITTI-360 validation with I2M+SeaBird, removing segmentation supervision (“No Loss”) reduces large-object AP from 8.71 to 4.86. Replacing Dice with other losses also degrades performance: Smooth-$L_2$6 gives 7.63, MSE gives 7.04, and cross-entropy gives 7.06 on the same large-object metric. The paper summarizes this as evidence that both the BEV formulation and Dice supervision are important, but Dice is the strongest of the tested objectives (Kumar et al., 2024).

Other ablations clarify the method’s inductive bias. Removing the segmentation head reduces detection quality. Replacing foreground-only segmentation with all-category segmentation causes a severe collapse in large-object performance, with Lrg dropping to 1.61 in one KITTI-360 ablation. Using only a single foreground class such as Car is also worse than segmenting all foreground detection categories. Sequential organization is better overall than parallel multitask design, even if a parallel row can show slightly higher Lrg in one table entry. Removing the BEV feature shortcut and supplying only the segmentation map to the detector also hurts performance, confirming that BEV occupancy is not sufficient by itself for full 3D box recovery (Kumar et al., 2024, Kumar, 27 Aug 2025).

The training schedule matters as well. The paper reports that segmentation pretraining followed by joint fine-tuning outperforms direct joint training and “detection then joint” variants. A denoiser based on MIRNet-v2 does not help and increases inference time, so it is not retained. In a nuScenes depth-estimation ablation on HoP+SeaBird, Dice again gives the best result, while the corresponding $L_2$7 version does not converge (Kumar et al., 2024).

SeaBird’s limitations are explicit. The theoretical analysis is simplified and does not cover all failure modes of real Mono3D systems. The method depends on a latent BEV representation, so it is not directly inserted into pure frontal-view detectors that never construct BEV features. The thesis also states that SeaBird does not fully solve the problem of generalization to large objects. Its contribution is narrower and more specific: it identifies large-object failure as partly a consequence of regression-loss sensitivity to depth noise, and shows that a sequential foreground BEV segmentation + Dice formulation can convert that insight into measurable improvements on KITTI-360 and nuScenes (Kumar, 27 Aug 2025).