ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects

AI summary

Overview

Research area: Computer vision — weakly supervised oriented object detection (OOD) in aerial imagery, specifically learning rotated bounding boxes from cheap horizontal bounding box annotations.

Technical level: Intermediate. The high-level idea is intuitive, but the paper assumes familiarity with oriented detection architectures, IoU-based regression losses, and self-supervised view-consistency training.

Scope: The paper introduces a weakly supervised OOD framework (ABBSPO) that stays accurate whether the ground-truth HBoxes are loose/coarse (C-HBoxes) or tight (T-HBoxes), and validates it on DIOR-R and DOTA-v1.0.

What This Paper Is About

Aerial objects like airplanes, ships, and overpasses need rotated bounding boxes (RBoxes) for precise localization, but RBox labels are expensive to annotate. Prior HBox-supervised methods (H2RBox, H2RBox-v2) compare ground-truth HBoxes against the minimum circumscribed rectangles of predicted RBoxes — a comparison that works when the labels are coarse HBoxes but degrades badly when they are tight HBoxes, such as in the DIOR dataset. The paper's goal is a single framework that handles both annotation styles by adaptively rescaling the ground-truth HBox per object and by adding a self-supervised angle signal derived from the left-right symmetry of aerial objects.

Key Contributions

1. First treatment of the T-HBox failure case in weakly supervised OOD. The authors state their work is the first to address the limitations of previous weakly supervised OOD methods when T-HBoxes are used as ground truth, proposing a method trainable with either T-HBoxes or C-HBoxes.

2. Adaptive Bounding Box Scaling (ABBS) module. It flexibly adjusts the ground-truth HBox for each object toward an appropriately scaled HBox, allowing part of the predicted RBox to lie outside the T-HBox and yielding more precise RBox size prediction.

3. Symmetric Prior Angle (SPA) loss. A self-supervised loss that leverages the inherent left-right symmetry of objects in aerial images to improve orientation prediction, targeting the failure case where the previous view-consistency losses give zero penalty despite consistently wrong angles across all three augmented views (original, rotated, flipped).

4. State-of-the-art weak-supervision results. The method claims significant gains over existing HBox-supervised OOD methods on aerial datasets.

Main Findings

• DIOR-R results: ABBSPO reaches 64.33 3-AP50 and 59.70 AP50, versus H2RBox at 51.43 / 54.57 and H2RBox-v2 at 55.23 / 56.67. This corresponds to reported improvements of 12.9%-point and 9.1%-point on 3-AP50 over H2RBox and H2RBox-v2 respectively, and 5.13%-point and 3.03%-point on overall AP50.

• 3-AP50 definition: On DIOR, 3-AP50 averages AP50 over three complex-shaped categories — 'airplane', 'expressway service area', and 'overpass'; on DOTA-v1.0 it averages over 'plane', 'swimming pool', and 'helicopter'.

• Comparison to fully supervised detectors: On DIOR-R, ABBSPO's 'airplane' AP is 69.5, above RetinaNet (59.8), FCOS (62.1), and Oriented R-CNN (63.0). The authors report margins of 6.5%-point to 11.7%-point over RBox-supervised methods on that category.

• DOTA-v1.0 results: On the validation split (458 images, not used for training), ABBSPO achieves 3-AP50 of 65.27 and AP50 of 69.26, beating H2RBox (61.30 / 66.07) by 3.19%-point and H2RBox-v2 (63.47 / 67.69) by 1.57%-point on AP50, and surpassing the FCOS baseline (68.6 AP50) by 0.66%-point.

• Evaluation caveat: The paper reports DOTA-v1.0 results on the validation set rather than the test set (937 images) because of non-responsiveness of the DOTA evaluation server.

• Ablation — both modules matter: Starting from the H2RBox-v2 baseline (55.23 3-AP50 / 56.67 AP50 on DIOR-R; 67.69 3-AP50 on DOTA), ABBS alone gives 62.13 / 58.35 and 68.59, SPA alone gives 58.77 / 58.99 and 69.16, and the combination gives 64.33 / 59.70 and 69.26.

• SPA applies only to symmetric classes: The model only applies the SPA loss to proposals for symmetric categories — 'airplane', 'ship', 'vehicle', and 'tennis court' — using symmetry priors from ground-truth class labels.

• Proposal sampling matters: Applying the SPA loss while sampling proposals from "others" (non-symmetric categories) collapses performance to 43.63 3-AP50 / 50.51 AP50 on DIOR-R, versus 64.33 / 59.70 with symmetric-category sampling. No SPA at all gives 61.67 / 58.93.

• Scale-range sensitivity: The best scale range in the ablation is Min 1.0, Max 1.5, Interval 0.1 (64.33 3-AP50 / 59.70 AP50 on DIOR-R). Narrower (0.9–1.1) and wider (1.0–2.0) ranges degrade results.

• Shape dependence: The ABBS module is reported as less effective for rectangular objects such as 'tennis court' and 'vehicle' because scaling is often unnecessary; gains concentrate on complex-shaped objects like the 'expressway service area'. For very small symmetric objects, the authors attribute weak SPA gains to insufficient pixels for reliably determining the symmetry axis.

• Orientation ambiguity handling: Round-shaped objects are given horizontal orientations to avoid confusion — DIOR-R: 'baseball field', 'chimney', 'golf field', 'stadium', 'storage tank', 'windmill'; DOTA-v1.0: 'baseball diamond', 'stadium', 'roundabout'.

• Supplemental datasets: SIMD and NWPU VHR-10 are listed among the evaluated datasets, but their results and details are deferred to the supplementary material and are not reported in the content available here.

Methodology in Plain English

The pipeline starts from the H2RBox-v2 framework and uses FCOS as the detector with a ResNet-50 backbone and FPN neck. Each image is fed in three views — original, rotated, and flipped — and the network predicts a rotated box per object (center, width, height, angle), plus class scores and a center-ness value. Classification and center-ness follow FCOS, and the angle head follows the PSC approach. The authors' own contributions sit in two branches:

Scale Learning Branch (SLB). Because a predicted RBox cannot be compared directly to an HBox label, it is converted into its minimum circumscribed rectangle (MCR). Instead of forcing that rectangle to match the ground-truth HBox, the module generates a set of scaled variants of the ground-truth HBox. The scale factors are evenly spaced across a predefined range and then adjusted by a piecewise-linear function of the predicted angle, so that objects at diagonal orientations get different scaling than axis-aligned ones. The variant with the lowest IoU loss across all proposals is selected as the target, and a small regularization term (weight α = 0.01) keeps the prediction tied to the unscaled HBox.

Orientation Learning Branch (OLB). Angles are supervised indirectly, through consistency between the rotated and flipped views using the snap losses from H2RBox-v2. On top of that, the SPA loss takes the top-k proposals ranked by classification plus localization score, projects each rotated box onto a fixed 50×50 grid, splits it along the box's symmetry axis into two parts, flips one part, and penalizes 1 minus the SSIM between the two halves. Objects are closer to correct when their two halves look alike, so the loss pushes the angle toward the true symmetry axis. The final angle loss combines the rotated and flipped losses with the SPA term using weights λr = 1.0, λf = 0.05, β = 0.6, and γ = 0.05, and the total loss adds regression, center-ness, and classification terms with all weighting factors set to 1.

All models were trained for 12 epochs on NVIDIA RTX3090 GPUs.

Why This Matters

Impact on research. The paper shows that a weak-supervision method does not have to be tied to one annotation style. By making C-HBox and T-HBox supervision both work, it removes a hidden assumption in earlier HBox-supervised work and opens existing tight-HBox datasets (like DIOR) to oriented detection without re-annotation. The SPA loss also offers a general trick — symmetry as free supervision — that is not specific to this architecture.

Real-world applications.

• Satellite and drone mapping of airports, ports, and road networks, where boxes must follow object orientation.
• Maritime and aviation monitoring, tracking ships and aircraft whose headings matter as much as their positions.
• Urban planning and infrastructure inspection, including overpasses, expressway service areas, and large facilities.
• Agricultural and environmental surveying, where fields, pools, and storage tanks appear in aerial views.

Industry relevance. The core selling point is cost: teams that already own HBox annotations can train oriented detectors without paying for rotated labels, and can reuse whichever annotation style their dataset happens to use. The reported parity with some fully RBox-supervised detectors on complex categories strengthens the case for weak supervision in production pipelines.

Future Directions

• Orientation-ambiguous objects. The paper sidesteps round objects by forcing horizontal labels. A principled way to handle them — rather than excluding or overriding them — remains open.
• Objects that break the symmetry assumption. SPA relies on left-right symmetry and is restricted to symmetric categories. Extending self-supervision to asymmetric objects, or handling occlusions, shadows, and asymmetric cargo arrangements robustly, is a natural next step.
• Scale-range generalization. Performance varies noticeably with the predefined scale range (best at 1.0–1.5, interval 0.1), suggesting the range is a hyperparameter that may need tuning per dataset or could be learned rather than fixed.
• Broader benchmarking. Full results on SIMD and NWPU VHR-10 are only referenced as being in the supplement, and the DOTA-v1.0 numbers are validation-set only because the evaluation server was unresponsive. Confirmation on the DOTA test set and additional datasets would strengthen the claims.

Target Audience

Researchers and engineers working on aerial or remote-sensing object detection, especially those interested in weakly supervised and label-efficient training. It is also relevant to practitioners who have horizontal bounding box annotations and want oriented detection without re-annotating, and to readers following self-supervised geometric priors and rotation-equivariant learning. Readers without prior exposure to oriented detection terminology will need to consult the referenced works (FCOS, H2RBox, H2RBox-v2, PSC) for full context.

Authors’ abstract

Weakly supervised oriented object detection (WS-OOD) has gained attention as a cost-effective alternative to fully supervised methods, providing both efficiency and high accuracy. Among weakly supervised approaches, horizontal bounding box (HBox)-supervised OOD stands out for its ability to directly leverage existing HBox annotations while achieving the highest accuracy under weak supervision settings. This paper introduces adaptive bounding box scaling and symmetry-prior-based orientation prediction, called ABBSPO, a framework for WS-OOD. Our ABBSPO addresses limitations of previous HBox-supervised OOD methods, which compare ground truth (GT) HBoxes directly with the minimum circumscribed rectangles of predicted RBoxes, often leading to inaccurate scale estimation. To overcome this, we propose: (i) Adaptive Bounding Box Scaling (ABBS), which appropriately scales GT HBoxes to optimize for the size of each predicted RBox, ensuring more accurate scale prediction; and (ii) a Symmetric Prior Angle (SPA) loss that exploits inherent symmetry of aerial objects for self-supervised learning, resolving issues in previous methods where learning collapses when predictions for all three augmented views (original, rotated, and flipped) are consistently incorrect. Extensive experimental results demonstrate that ABBSPO achieves state-of-the-art performance, outperforming existing methods.