🤖 machine learning

Hausdorff Distance Matching with Adaptive Query Denoising for Rotated Detection Transformer

This paper proposes a Rotated Detection Transformer that improves oriented object detection by introducing a Hausdorff distance-based cost for more accurate bipartite matching and an adaptive query denoising method to overcome the limitations of static denoising, achieving significant performance gains on multiple benchmarks.

Original authors: Hakjin Lee, MinKi Song, Jamyoung Koo, Junghoon Seo

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hakjin Lee, MinKi Song, Jamyoung Koo, Junghoon Seo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize objects in a huge, messy aerial image of a city. Some objects, like cars, are easy to find because they are usually straight and box-shaped. Other objects, such as ships in a harbor or airplanes on a runway, are often tilted at strange angles. To find these, the robot must draw a tilted box around them.

This work introduces a new robot brain (called RHINO) that finds these tilted objects significantly better than previous models. The authors identified two main reasons why older robots struggled and fixed them with two clever tricks.

The Problem: The "Square" Confusion and the "Noisy" Teacher

1. The "Square" Confusion (The Hausdorff Distance Correction)
Imagine a game where you must match a series of red stickers (the robot's guesses) with blue stickers (the actual objects).

The old way: The old robot used a simple ruler to measure the distance between the center of the red sticker and that of the blue one. Because tilted objects can look like squares from certain angles, the ruler got confused. It sometimes said: "Hey, this red sticker is far away, but it has the same angle as the blue one, so it's a hit!" This caused the robot to draw two boxes for the same object: one good one and one poor one with low confidence.
The new way (RHINO): The authors realized that instead of looking only at the center, they should consider the entire shape of the box. They used a mathematical tool called Hausdorff distance. Imagine this as measuring the distance between the edges of the shapes, not just between their centers. It is like checking whether the corners of the red sticker actually match the corners of the blue one. This prevented the robot from being confused by square-like shapes and eliminated these duplicate, useless boxes.

2. The "Noisy" Teacher (Adaptive Query Denoising)
To teach the robot faster, the old method used a technique called "Query Denoising." Imagine a teacher trying to teach a student by giving them a smeared, distorted version of the correct answer and asking them to clean it up.

The problem: At the beginning of training, the robot is very poor, so the teacher's smeared notes are actually helpful. But as soon as the robot becomes smarter and makes perfect predictions, the teacher continues to give it the same smeared, distorted notes. The robot gets confused: "Wait, I know the answer is perfect, but the teacher is telling me to fix this smeared version. Should I follow the teacher or my own brain?" This actually slowed down the robot's learning in later stages.
The new way (RHINO): The authors made the teacher adaptive. They added a "filter" that checks the robot's current skill level.
- If the robot is a beginner, the filter lets the smeared notes through.
- If the robot is an expert and its own predictions are already better than the smeared notes, the filter throws away the smeared notes. It tells the robot: "You no longer need to fix this junk; trust your own perfect answer." This keeps training focused and efficient.

The Results

The authors tested this new robot (RHINO) on several famous aerial image datasets (such as DOTA and DIOR-R).

The score: Compared to the previous best models using the same basic hardware (a ResNet-50 backbone), RHINO achieved significantly higher values. It improved accuracy by about 4 to 5 points on a scale where higher values are better.
The comparison: It beat other top models, even those using much more powerful and complex computer brains (backbones).

In Brief

The work says: "We built a better detector for tilted objects by fixing two things:

We changed how the robot measures distances so it stops drawing duplicate boxes for square-like objects.
We made the training process smarter so it stops listening to 'noisy' examples once the robot has already learned the correct answer."

The result is a model that finds tilted objects (such as ships and airplanes) more accurately and with fewer errors than before.

Technical Summary: Hausdorff Distance Matching with Adaptive Query Denoising for Rotated Detection Transformer

Problem Statement
While Detection Transformers (DETR) have established new benchmarks in horizontal object detection, their application to rotated object detection (using oriented bounding boxes) lags behind specialized oriented detectors. The authors identify two primary bottlenecks hindering the performance of rotated DETR models:

The "quadratic" problem in bipartite matching: Standard bipartite matching in DETR relies on L1 distance for localization costs. In rotated detection, this metric causes discontinuities at boundaries and "quadratic" problems where the model assigns ground-truth objects to predictions that share the same angle but are spatially distant or have low Intersection over Union (IoU). This leads to duplicated predictions with low confidence, contradicting the core DETR principle of one-to-one matching.
Limitations of static query denoising: Existing query denoising methods (e.g., in DINO) introduce noise into ground-truth data to accelerate early training. However, once the model improves, its predictions may become more accurate than the noisy ground-truth data. Static denoising forces the model to continue learning from these inferior, noisy targets, hindering performance in later training phases.

Methodology
The authors propose RHINO (Rotated detection transformer with Hausdorff distance matching), a DINO-based baseline extended with two novel components:

Hausdorff Distance for Bipartite Matching:
To solve the quadratic problem, the authors replace standard L1 distance costs with Hausdorff distance.
- Mechanism: Hausdorff distance calculates the maximum distance between the point sets defining two bounding boxes (specifically using corner points or edge midpoints).
- Effect: Unlike L1, which overemphasizes angle differences even after normalization, Hausdorff distance accounts for axis alignment and spatial overlap of the boxes. This ensures ground-truth objects are matched to predictions with higher geometric consistency, effectively eliminating duplicated low-confidence predictions.
- Implementation: The model extends the standard 4D box representation to 5D coordinates $[x, y, w, h, \theta]$ and maps the angle $\theta$ from $[0, 180^\circ]$ to $[0, 1]$ to seamlessly integrate with existing DETR components such as iterative refinement and deformable attention.
Adaptive Query Denoising:
To address the problem of learning from noisy targets in later training phases, the authors introduce an adaptive filtering mechanism.
- Mechanism: Instead of blindly using all positive noisy queries for supervision, the method employs bipartite matching between the set of refined noisy queries and the model's current predictions.
- Filtering: Only noisy queries successfully matched to their original ground-truth source (i.e., those the model has not yet "surpassed" or that still represent the best match) are retained for the denoising loss. Queries that no longer represent the optimal match are filtered out or treated as background.
- Enhanced Variant: An "Enhanced Adaptive Query Denoising" method (AQD*) is proposed, which retains a regression loss for filtered-out queries to further guide localization refinement, even when they are not selected for the primary classification objective.

Main Contributions

Robust Baseline: The authors design a simple, robust DETR baseline for rotated detection that integrates state-of-the-art improvements (iterative refinement, dynamic anchor boxes, contrastive denoising) without requiring complex, hand-crafted modules or specialized loss functions beyond the proposed modifications.
Hausdorff Distance Costs: Introducing Hausdorff distance as matching costs successfully reduces the quadratic problem and duplicated predictions, leading to more stable training and higher performance.
Adaptive Query Denoising: The proposal of a dynamic filtering strategy that selectively eliminates harmful noisy queries once model accuracy improves prevents the model from reverting to noisy targets in later training phases.

Experimental Results
The model was evaluated on four standard datasets for rotated object detection: DOTA-v1.0, DOTA-v1.5, DOTA-v2.0, and DIOR-R, under single-scale training and testing settings.

Performance Gains: Compared to models with a ResNet-50 backbone, RHINO achieved significant improvements:
- DOTA-v2.0: +4.18 AP50.
- DOTA-v1.5: +4.59 AP50.
- DIOR-R: +4.99 AP50.
State-of-the-Art: RHINO set new benchmarks on all tested datasets. For example, RHINO with a Swin-Tiny backbone achieved 79.42 AP50 on DOTA-v1.0, outperforming previous DETR-based and convolution-based detectors.
Precision Metrics: The model also demonstrated superior performance on AP75 (a stricter metric for orientation accuracy), indicating that Hausdorff distance effectively improves angular precision.
Ablation Studies: Experiments confirmed that replacing L1 with Hausdorff distance reduces duplicated predictions and that adaptive denoising outperforms static denoising with fixed noise levels.

Significance
The work claims that RHINO demonstrates that query-based detectors can match or exceed the performance of specialized oriented object detectors without relying on complex, domain-specific architectural changes. By addressing fundamental mismatches in bipartite matching and the limitations of static denoising in the context of rotated objects, the study provides a strong, DINO-based baseline for future research in oriented object detection. The work underscores the potential to adapt general DETR architectures to rotated detection through careful metric selection and adaptive training strategies.

The Problem: The "Square" Confusion and the "Noisy" Teacher

The Results

In Brief

More like this