SPWOOD: Sparse Partial Weakly-Supervised Oriented Object Detection

Conference: ICLR 2026 arXiv: 2602.03634 Code: N/A Area: Object Detection / Remote Sensing Keywords: Oriented object detection, weak supervision, sparse annotation, semi-supervised learning, remote sensing

TL;DR

This paper proposes the SPWOOD framework to jointly address sparse and weak annotation (HBox/Point) in oriented object detection. Through a Self-Adaptive Oriented Detector (SAOD) and a spatial layout learning strategy, SPWOOD achieves near-fully-supervised performance on the DOTA benchmark under a mixed annotation setting (RBox:HBox:Point = 1:1:1).

Background & Motivation

Background: Oriented object detection (OOD) is critical in remote sensing and related domains. However, annotating precise rotated bounding boxes (RBoxes) — requiring center coordinates, width, height, and rotation angle — incurs prohibitively high labeling costs.

Limitations of Prior Work: Existing methods for reducing annotation cost either handle only weak annotations (e.g., replacing RBoxes with horizontal boxes HBox or point annotations) or only sparse annotations (i.e., a subset of instances are labeled), but in practice both challenges co-exist.

Key Challenge: Sparse annotations (not all instances are labeled) and weak annotations (labeled but imprecise) each introduce severe training signal deficiency; their combination exacerbates the problem — unlabeled instances may be treated as negatives, while weak annotations may mislead angle estimation.

Goal: To train high-quality oriented object detectors under an extremely low-cost setting where both sparse and weak annotations are present simultaneously.

Key Insight: Design a unified framework that learns from three annotation types of varying quality (RBox, HBox, Point) while mining unlabeled instances via self-training.

Core Idea: A self-adaptive oriented detector unifies precise, weak, and unlabeled training signals; angular consistency constraints and spatial layout learning recover rotation information from weak annotations.

Method

Overall Architecture

SPWOOD builds upon a teacher–student self-training framework and introduces three key components: (1) SAOD, a baseline detector that handles RBoxes and pseudo-RBoxes recovered from weak annotations; (2) an angle learning module that exploits geometric consistency under image augmentation to supervise rotation estimation; and (3) a spatial layout learning strategy that recovers scale information for point-annotated instances via Voronoi–watershed analysis.

Key Designs

1. Self-Adaptive Oriented Detector (SAOD):

   • Function: A unified detection baseline that handles annotations of heterogeneous quality.
   • Mechanism: Standard rotated detection losses are applied to RBox annotations. For HBox annotations, feasible rotation ranges are inferred from the horizontal box to generate pseudo-RBoxes for training. For Point annotations, only center location is used, with scale recovered via spatial layout learning. The teacher model, trained on strongly annotated instances, generates pseudo-labels for unlabeled regions.
   • Design Motivation: Different annotation types provide complementary information along different dimensions, necessitating differentiated rather than uniform treatment.

2. Angle Learning:

   • Function: Recover rotation angles from HBox and Point annotations that contain no explicit angle information.
   • Mechanism: Geometric consistency under image augmentation is exploited — given a rotation augmentation of angle \(\mathcal{R}\), the predicted angles must satisfy \(\theta_{rot} = \theta + \mathcal{R}\). The loss is defined as: \(\mathcal{L}_{Ang}^s = \text{SmoothL1}(\theta_{flp} - \theta_{flp}^{gt}) + \text{SmoothL1}(\theta_{rot} - \theta - \mathcal{R})\)
   • Design Motivation: Rotation angle is the hardest attribute to recover from weak annotations, yet geometric transformations provide natural self-supervised signals.

3. Spatial Layout Learning:

   • Function: Recover object width and height from point-only annotations.
   • Mechanism: A Voronoi diagram is constructed from the set of point annotations to partition the image into per-point regions. Within each region, a watershed algorithm segments pixels based on appearance similarity to obtain pixel-level assignments. The watershed results are rotation-aligned to yield regression targets for width and height. The Voronoi–watershed loss is: \(\mathcal{L}_W^s = L_{GWD}(\text{pred}, \text{watershed\_target})\)
   • Design Motivation: Voronoi diagrams naturally separate adjacent objects, and the watershed algorithm provides appearance-based scale estimation within each region.

Loss & Training

The total loss comprises: detection loss (standard rotated box regression + classification) + angle consistency loss \(\mathcal{L}_{Ang}\) + spatial layout loss \(\mathcal{L}_W\) + Gaussian overlap loss \(\mathcal{L}_O\) (penalizing overlapping predictions). The teacher–student framework is updated via Exponential Moving Average (EMA).

Key Experimental Results

Main Results

Method	Type	10% Sparse · 10% Partial	20% · 20%	30% · 20%
H2RBox-v2	Weak supervision (HBox)	30.6	42.7	49.2
MCL	Semi-supervised (RBox)	31.7	44.5	47.8
PWOOD	Partial weak supervision (RBox)	38.0	51.9	55.2
RSST	Sparse supervision (RBox)	43.4	52.3	56.6
SPWOOD (RBox)	Sparse + Weak	48.5	57.8	60.3
SPWOOD (HBox)	Sparse + Weak	45.5	54.0	56.5
SPWOOD (R:H:P=1:1:1)	Mixed	42.4	53.0	54.8

Ablation Study

Configuration	mAP (10%·10%)	Notes
SPWOOD (full)	48.5	All components
w/o Angle Learning	~43	Inaccurate angles under weak annotation
w/o Spatial Layout	~44	Poor scale recovery for point annotations
w/o Teacher–Student	~40	Unlabeled instances unexploited

Key Findings

• SPWOOD (RBox) consistently outperforms all baselines across annotation ratios, with gains up to 5+ mAP.
• The mixed annotation setting (R:H:P=1:1:1) achieves performance close to fully sparse RBox supervision.
• Angle learning contributes most significantly in weak annotation scenarios.
• Spatial layout learning becomes increasingly critical under extremely sparse settings.

Highlights & Insights

• Unified framework for heterogeneous annotation types: SPWOOD elegantly integrates three annotation sources, each providing information of different quality and granularity.
• Effective exploitation of geometric consistency: Rotation angle is self-supervised via augmentation-induced angle constraints, requiring no explicit angle labels.

Limitations & Future Work

• The Voronoi–watershed approach may degrade in highly dense object scenes.
• The angle learning module assumes that augmentation transformations are known, limiting applicability to scenarios with unknown viewpoint variations.
• Evaluation is conducted exclusively on the DOTA remote sensing benchmark; generalization to natural image oriented detection remains unexplored.

Related Work & Insights

• vs. Point2RBox: Recovers rotated boxes from point annotations only, without addressing sparse annotation.
• vs. PWOOD: Handles partial weak supervision but not sparsity — it assumes all instances carry at least weak annotations.

Rating

• Novelty: ⭐⭐⭐⭐ First unified treatment of sparse + weak annotation for oriented object detection.
• Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive comparisons across multiple annotation ratios and baselines.
• Writing Quality: ⭐⭐⭐ Method descriptions are clear, though the notation is dense.
• Value: ⭐⭐⭐⭐ Directly applicable to low-cost remote sensing detection pipelines.

Related Papers

• [CVPR 2026] Fourier Angle Alignment for Oriented Object Detection in Remote Sensing
• [ICLR 2026] Long-Context Generalization with Sparse Attention
• [AAAI 2026] TubeRMC: Tube-conditioned Reconstruction with Mutual Constraints for Weakly-supervised Spatio-Temporal Video Grounding
• [ICLR 2026] CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection
• [ICLR 2026] FSOD-VFM: Few-Shot Object Detection with Vision Foundation Models and Graph Diffusion