bounding box postprocessing

Existing IoU-based bounding box regression losses suffer from gradient vanishing in non-overlapping cases and rely on handcrafted geometric penalty terms, leading to suboptimal small-object optimization and anomalous box expansion. To address these issues, we propose InterpIoU, which replaces conventional geometric penalties with a differentiable interpolated-box IoU—resolving gradient discontinuity and misalignment at the source. We further introduce Dynamic InterpIoU, which adaptively adjusts the interpolation coefficient based on object scale and spatial distribution. Our method requires no additional hyperparameters and is plug-and-play compatible with mainstream detectors. Extensive experiments on COCO, VisDrone, and PASCAL VOC demonstrate consistent superiority over state-of-the-art IoU-based losses: notably, small-object AP improves by +2.1–3.4 points, validating both effectiveness and generalizability across diverse detection scenarios.

Small-object detection suffers from significant performance degradation due to inaccurate localization and unstable gradients. This paper identifies that conventional regression-based bounding box localization induces distorted gradients for small objects. To address this, we reformulate bounding box localization as a grid-based classification task—the first such approach—and propose a confidence-driven localization framework. Our method employs two-hot label encoding, confidence distribution prediction, cross-entropy localization loss, and an entropy-based uncertainty loss to jointly correct gradient flow and suppress localization uncertainty. Evaluated on mainstream detectors—including YOLOv8 and RT-DETR—and across three major benchmarks (COCO, VisDrone, and AI-TOD), our approach achieves state-of-the-art performance, notably improving AP for small objects. Moreover, it demonstrates strong generalization across diverse annotation protocols and high-resolution imagery.

This paper addresses the challenging fitting problem in constructive solid geometry (CSG) modeling for scene geometric parsing—characterized by unknown numbers of primitives, strong coupling between positive and negative primitives, and highly non-convex optimization. We propose a multi-regression ensemble framework integrating learned initialization, gradient-based geometric optimization, and Boolean composition modeling. Crucially, we introduce explicit representation of *negative convex primitives* to model voids and occlusions—a first in CSG-based reconstruction. Furthermore, we design a “refine-then-select” ensemble strategy that breaks from conventional paradigms. On standard benchmarks, our method significantly reduces depth and surface normal prediction errors; over 70% of images benefit from negative primitive modeling. The ensemble strategy substantially outperforms baselines, empirically confirming the problem’s severe non-convexity and validating the effectiveness of our approach.

This work addresses the challenge of formally verifying the robustness of object detection models against input perturbations. We propose the first abstract interpretation framework specifically designed for IoU-based robustness certification, extending Interval Bound Propagation (IBP) to rigorously reason about IoU—a non-differentiable, geometric metric—under input uncertainty. Our method precisely models the interval propagation of bounding box coordinates through the IoU computation graph, enabling tighter and more stable bounds than existing IBP-based baselines. Experiments on runway detection and handwritten digit recognition demonstrate both empirical effectiveness and strong generalization across diverse detection tasks. The framework is implemented as open-source code, fully compatible with mainstream abstract interpretation verification toolkits.

Existing interpretability methods for object detection predominantly rely on single-pixel attribution, failing to capture the joint influence of multi-pixel collaborations on both bounding box localization and class prediction—thus overlooking compositional cues or introducing spurious correlations. To address this, we introduce Shapley interaction values to object detection interpretation for the first time, proposing the first end-to-end differentiable framework that explicitly models high-order cooperative effects among pixel groups. Our approach jointly models feature-space perturbations and detection output sensitivity to simultaneously quantify individual pixel contributions and higher-order interactions. Extensive experiments on COCO and other benchmarks demonstrate that our method significantly outperforms state-of-the-art interpretability baselines. Both qualitative visualizations and quantitative metrics confirm its superior ability to localize discriminative visual regions accurately. The source code will be made publicly available.

This work addresses the limitation of existing IoU-based loss functions, which employ fixed geometric penalty terms and thus struggle to adapt to the dynamic optimization requirements of bounding box regression across different training stages. To overcome this, the authors propose MoEIoU, the first IoU loss grounded in a Mixture-of-Experts (MoE) mechanism, formulating regression as a mixture-of-experts problem. MoEIoU leverages the log-sum-exp function to smoothly and adaptively aggregate errors in overlap, center alignment, and aspect ratio, while integrating a curriculum learning strategy to dynamically adjust the weights of these error components during training. Extensive experiments on PASCAL VOC, HRIPCB, and MS COCO demonstrate that MoEIoU consistently outperforms current state-of-the-art losses, yielding faster convergence, improved localization accuracy, and notable performance gains across multiple YOLO architectures.

This work addresses the limitations of existing oriented bounding box regression methods, which rely solely on geometric information while neglecting semantic cues and suffer from unstable gradients during orientation optimization, leading to convergence oscillations. To overcome these issues, the authors propose Fractional-order Semantic IoU loss (FrSIoU), which, for the first time, integrates semantic similarity into IoU-based optimization. By leveraging fractional-order calculus to accumulate historical IoU states, FrSIoU establishes an adaptive gradient gating mechanism that unifies semantic and geometric modeling. Extensive experiments demonstrate consistent performance gains across diverse bounding box representations and visual detection tasks, effectively mitigating gradient oscillations and validating the method’s efficacy and generalization capability.

Existing image manipulation localization methods rely primarily on 2D cues and suffer significant performance degradation when tampered regions are seamlessly blended with the background. This work proposes a geometry-aware localization framework that, for the first time, incorporates 3D geometric cues—such as depth and surface normals derived from monocular reconstruction—into the task. By assessing the reliability of these 3D cues, the method employs a multi-scale fusion mechanism to selectively integrate them with RGB features. The proposed approach achieves notably improved localization accuracy with minimal additional computational overhead, demonstrating that trustworthy 3D geometric information effectively complements conventional 2D forensic cues.

This work addresses the challenge of monocular 3D object detection in real-world scenarios where camera intrinsics are unknown, making accurate recovery of 3D geometry from the image plane difficult. To this end, we propose MoCA3D, a category-agnostic monocular 3D detection model that eliminates the need for camera intrinsics at inference time by directly predicting 3D bounding box corners and their depths densely in the image plane via corner heatmaps and depth maps. Departing from conventional paradigms that lift 2D RoIs to 3D boxes, our method introduces pixel-aligned geometric constraints, enabling high-fidelity, intrinsic-agnostic 3D geometry prediction for the first time. We further propose the PAG metric to evaluate consistency between predicted corners and depths, achieving a 22.8% improvement on this metric while matching state-of-the-art 3D IoU performance, with model parameters reduced by up to 57×, significantly enhancing applicability in settings lacking camera calibration.

This work addresses the challenge of robustly lifting 2D object detections to metric 3D bounding boxes in open-world scenarios where dense depth maps and 3D annotations are scarce. To this end, we propose BoxerNet, a Transformer-based architecture that integrates open-vocabulary 2D detectors (e.g., OWLv2, DETiC), multi-view geometric constraints, and optional depth cues—either sparse or dense. A median-depth block enables effective handling of sparse depth inputs, while aleatoric uncertainty modeling enhances regression robustness. By combining geometric filtering with multi-view fusion, BoxerNet produces globally consistent, de-duplicated 3D boxes. Trained on over 1.2 million unique 3D bounding boxes, our model significantly reduces reliance on costly 3D annotations, achieving an mAP of 0.532 in ego-centric settings without dense depth—substantially outperforming CuTR (0.010)—and 0.412 on CA-1M, surpassing CuTR’s 0.250.