TD-RD: A Top-Down Benchmark with Real-Time Framework for Road Damage Detection

Xi Xiao, Zhengji Li, Wentao Wang, Jiacheng Xie, Houjie Lin, Swalpa Kumar Roy, Tianyang Wang, Min Xu

Road maintenance is a high-stakes game of "find the needle in the haystack," where the needles are potholes and cracks spread across thousands of miles of asphalt. While computer vision has made huge strides in detecting these hazards, most existing models look at the road through the "eyes" of a dashcam. This oblique perspective can distort the size and shape of damage.

A new paper, "TD-RD: A Top-Down Benchmark with Real-Time Framework for Road Damage Detection," proposes a shift in perspective—literally. By moving to a top-down view and introducing a specialized deep learning architecture, the authors aim to make road inspection more accurate and efficient.

TL;DR

The authors introduce TD-RD, the first high-resolution, top-down dataset for road damage, and TD-YOLOv10, a real-time detection framework. Their model achieves a high balance of accuracy (88.1% mAP) and speed (200 FPS), significantly outperforming standard YOLO variants on pavement-specific tasks.

What Problem Does This Paper Address?

Most road damage datasets (like CRDDC or CNRDD) rely on images captured from vehicle dashboards. While convenient, this perspective suffers from perspective distortion, where damage further away appears smaller and more compressed.

To truly model road surface conditions for engineering and maintenance, a top-down (nadir) view is superior. It provides a "map-like" representation that makes it easier to measure the actual area of a pothole or the length of a crack. However, top-down data is scarce, and standard object detectors often struggle with the fine-grained, multi-scale nature of pavement distress (e.g., a tiny crack vs. a large repair patch).

Methods

The researchers developed TD-YOLOv10, a real-time framework optimized for the unique textures of asphalt. The architecture introduces two primary innovations:

  1. MAPSE (Multi-Scale Attention with Positional Squeeze-and-Excitation): This module is integrated into the model's backbone. It uses multi-head attention with varying "dilation rates" (r=1, 2, 4). This allows the model to "see" at different scales simultaneously—essential for detecting both hairline cracks and massive potholes. It also uses a squeeze-and-excitation mechanism to recalibrate feature maps, focusing the model's "attention" on the most relevant parts of the image.
  2. VGAU (Visual Global Attention Upsampling): Standard upsampling often loses semantic context. VGAU uses global average pooling to incorporate global information during the upsampling process in the model's "neck," helping the detector distinguish between actual damage and harmless road textures or shadows.
  3. Efficiency: To maintain real-time speeds, the model utilizes Depthwise Separable Convolutions (DSC), which drastically reduce the number of parameters and computational load without a significant hit to accuracy.

Datasets

The study utilizes three key datasets:

  • TD-RD (New): The authors' primary contribution. It contains 7,088 high-resolution images (3840 × 2160 pixels) with 12,882 annotated instances. It focuses on three balanced categories: cracks, potholes, and patches (repairs).
  • CNRDD & CRDDC’22: Existing public datasets used to benchmark the model's robustness across different camera angles and road conditions.

Key Insights and Ideas

  • The Power of Perspective: The top-down view provides a "holistic approach" to capturing road intricacies. The results suggest that specialized benchmarks like TD-RD are necessary because models trained on dashcam data don't always generalize well to the views used by professional road inspection vehicles.
  • Attention is Mandatory for Pavement: Because road damage is often "low-contrast" (grey cracks on grey asphalt), global context is vital. The MAPSE module's ability to capture long-range dependencies helps the model understand that a series of small dark pixels is actually a continuous longitudinal crack.
  • Balanced Data Matters: Unlike many datasets that are heavily skewed toward cracks, TD-RD is curated to be balanced across cracks, potholes, and repairs, ensuring the model doesn't become biased toward the most common type of damage.

Key Results

The TD-YOLOv10 model demonstrates a "sweet spot" between speed and precision:

  • Accuracy on TD-RD: Achieved 88.1% mAP (Mean Average Precision) and 88.5% Precision, outperforming YOLOv8-s (+3.0% mAP) and the state-of-the-art RT-DETR (+0.4% mAP).
  • Real-Time Performance: The model sustains a frame rate of 200 FPS on an Nvidia A100 GPU, making it highly suitable for deployment on high-speed inspection vehicles.
  • Robustness: On the challenging CRDDC’22 dataset, it achieved 48.0% mAP, surpassing other real-time models like PP-PicoDet and YOLOv10-s.

Limitations and Future Work

While the paper presents a robust framework, a few areas remain for exploration:

  • Real-World Deployment: The 200 FPS benchmark was performed on a high-end A100 GPU; performance on edge devices (like those actually mounted on maintenance trucks) wasn't explicitly detailed.
  • Weather and Lighting: The paper focuses on high-quality images. Future work could explore how the top-down perspective handles adverse conditions like rain, snow, or night-time infrared captures.
  • Multi-View Fusion: The authors suggest that combining their top-down data with existing multi-view sources could lead to even more comprehensive road modeling.

Reference: Xiao, X., et al. "TD-RD: A Top-Down Benchmark with Real-Time Framework for Road Damage Detection." (2024).

Summary updated

Abstract

Object detection has witnessed remarkable advancements over the past decade, largely driven by breakthroughs in deep learning and the proliferation of large-scale datasets. However, the domain of road damage detection remains relatively underexplored, despite its critical significance for applications such as infrastructure maintenance and road safety. This paper addresses this gap by introducing a novel top-down benchmark that offers a complementary perspective to existing datasets, specifically tailored for road damage detection. Our proposed Top-Down Road Damage Detection Dataset (TD-RD) includes three primary categories of road damage—cracks, potholes, and patches—captured from an top-down viewpoint. The dataset consists of 7,088 high-resolution images, encompassing 12,882 annotated instances of road damage. Additionally, we present a novel real-time object detection framework, TD-YOLOV10, designed to handle the unique challenges posed by the TD-RD dataset. Comparative studies with state-of-the-art models demonstrate competitive baseline results. By releasing TD-RD, we aim to accelerate research in this crucial area. A sample of the dataset will be made publicly available upon the paper’s acceptance.
BibTeX
@inproceedings{xiao2025td,
  author = {Xi Xiao and Zhengji Li and Wentao Wang and Jiacheng Xie and Houjie Lin and Swalpa Kumar Roy and Tianyang Wang and Min Xu},
  title = {{TD-RD: A Top-Down Benchmark with Real-Time Framework for Road Damage Detection}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=vygpHQu3Hk},
}

Loading related content...

Figure