Gradient-Guided Orientation Learning for Weakly Supervised Oriented Object Detection

Junyong Jang, Juheon Park, Minyoung Jeon, Jaehyup LeearXiv 2026 · arXiv

In the world of remote sensing—think satellite imagery of harbors, airports, and highways—objects rarely align neatly with the axes of a digital image. Detecting a ship or a runway requires Oriented Object Detection (OOD): drawing rotated bounding boxes that capture both location and heading.

The catch? Annotating these rotated boxes is incredibly labor-intensive. While researchers have turned to Weakly Supervised (WS-OOD) methods that learn from simpler horizontal boxes (HBoxes), these models often struggle to "guess" the correct angle accurately. A new paper, Gradient-Guided Orientation Learning for Weakly Supervised Oriented Object Detection, proposes a clever way to let the images teach themselves the right orientation.

TL;DR

GradOrient improves weakly supervised oriented object detection by extracting orientation cues directly from image gradients to create "pseudo-angle" labels, eliminating the need for expensive manual rotated-box annotations. The authors also introduce Orientation-Aware AP (OAP), a more rigorous evaluation metric that penalizes angular errors which standard metrics often overlook.

What Problem Does This Paper Address?

Current weakly supervised methods usually rely on "geometric consistency." For example, if you rotate an image 30 degrees, the model’s predicted box should also rotate 30 degrees. While logical, this is an indirect signal. In cluttered scenes or images with shadows and low contrast, this consistency can be satisfied by multiple incorrect angles, leading to unstable training.

Furthermore, the standard metric for success, Average Precision (AP), is surprisingly "angle-blind." As shown in Figure 1(a), two predicted boxes can have nearly identical Intersection-over-Union (IoU) scores despite one having a massive 16-degree heading error. In applications like maritime monitoring or aircraft pose estimation, that error is the difference between a success and a failure.

Figure 1: (a) Shows how standard IoU fails to capture large angle errors. (b) & (c) show how GradOrient produces tighter, more accurate alignments.

Methods

The authors introduce GradOrient, which consists of two main pillars: a new supervision strategy and a better way to measure results.

1. The Pseudo-Angle Generator

Instead of relying solely on geometric tricks, GradOrient looks at the image gradients. Most objects in remote sensing (ships, planes, buildings) have strong, elongated edges that align with their physical orientation.

  • Histogram of Gradients: Within a horizontal box, the model computes local image gradients and builds a magnitude-weighted histogram.
  • Dominant Orientation: The peak of this histogram represents the "dominant" direction of the object's edges.
  • Confidence Gating: Not all gradients are reliable. The system calculates a confidence score based on the histogram's entropy. If the gradients are scattered (low confidence) or the pseudo-angle is too far from the current prediction, the signal is ignored to prevent noisy training.
Figure 2: The GradOrient pipeline. Gradients are extracted from HBox regions, converted to edge tangents, and used to supervise the detection head.

2. Orientation-Aware AP (OAP)

To fix the "angle-blindness" of standard evaluation, the authors propose OAP. They modify the IoU calculation by multiplying it by an angular penalty weight ($W$). They specifically chose a square-root penalty (see Figure 3), which is much harsher on small angular discrepancies than linear or cosine-based penalties. This ensures that a "True Positive" actually reflects a box that is well-aligned in both space and rotation.

Figure 3: Comparison of penalty weights. The square-root penalty (red) drops sharply for small angle gaps, enforcing stricter orientation accuracy.

Datasets

The researchers validated their approach on two gold-standard benchmarks for aerial detection:

  • DOTA-v1.0: 2,806 high-resolution images containing over 188,000 instances across 15 categories.
  • DIOR-R: A massive remote sensing dataset with 23,463 images and 192,518 instances across 20 categories.

Key Insights and Ideas

  • The Image is the Label: The core "aha!" moment is that the orientation information is already present in the raw pixels. By using gradients as a prior, the model gains a direct, image-intrinsic signal for angle learning without needing a single extra human-annotated rotated box.
  • Plug-and-Play Supervision: GradOrient doesn't require a specific architecture. It acts as an auxiliary loss that can be plugged into existing frameworks like FCOS or H2RBox.
  • Metric Reform: By showing that standard AP masks significant errors, the paper advocates for a shift in how the OOD community measures progress, prioritizing angular precision alongside spatial overlap.

Key Results

GradOrient consistently outperformed existing state-of-the-art weakly supervised methods:

  • Accuracy Boost: On the DIOR-R dataset, it achieved 61.50 AP₅₀, outperforming the strong ABBSPO baseline by 1.4 points.
  • Finer Alignment: The "AngleGap" (average angular error) dropped from 5.02° to 4.16° on DIOR-R.
  • Qualitative Superiority: As seen in Figure 4, GradOrient produces boxes that are visibly more consistent with object directions, especially for densely packed or highly elongated objects where previous methods often "tilted" incorrectly.
Figure 4: Visual results on DIOR-R and DOTA-v1.0. GradOrient (right) shows significantly better alignment with the actual axes of ships and planes compared to baselines.

Limitations and Future Work

The authors note that pseudo-angle estimation can be tricked by "zebra-like" textures—internal patterns (like stripes on a ship's deck) that don't align with the object's main axis. Similarly, highly cluttered backgrounds can introduce misleading gradients.

While their confidence-gating mechanism mitigates this, future work could involve more sophisticated filtering or using deep features instead of raw gradients to identify the "true" primary axis of an object. Additionally, extending this gradient-guided logic to point-supervised detection (where even the HBox is missing) remains an open challenge.

Summary updated

Abstract

Weakly supervised oriented object detection (WS-OOD) aims to reduce the annotation cost of rotated bounding boxes by leveraging cheaper annotations such as horizontal boxes or points. However, these weak annotations provide no explicit orientation information, forcing existing methods to infer object angles through indirect geometric consistency or equivariant constraints. Such indirect supervision often becomes ambiguous and unstable under real-world appearance variations, including illumination changes, color shifts, shadows, and background clutter. In this paper, we present a simple yet effective gradient-guided orientation learning strategy for WS-OOD. Motivated by the observation that the input image itself contains useful orientation priors, our method extracts dominant gradient-based orientation cues from object regions and converts them into pseudo-angle labels for auxiliary supervision, enabling the detector to learn orientation-aware representations without requiring rotated bounding box annotations or extra learnable parameters. Furthermore, since AP can mask substantial angular errors when spatial IoU remains high, we introduce an orientation-aware evaluation metric that incorporates angular discrepancies into the standard AP protocol. Extensive experiments on weakly supervised oriented detection benchmarks demonstrate that the proposed pseudo-angle supervision improves both detection performance and orientation estimation quality, while the proposed metric provides a more faithful assessment of angle prediction than conventional AP alone.
BibTeX
@misc{jang2026gradient,
  author = {Junyong Jang and Juheon Park and Minyoung Jeon and Jaehyup Lee},
  title = {{Gradient-Guided Orientation Learning for Weakly Supervised Oriented Object Detection}},
  year = {2026},
  eprint = {2611.04616},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.04616},
}
Figure