RoLID-11K: A Dashcam Dataset for Small-Object Roadside Litter Detection

TL;DR

RoLID-11K introduces the first large-scale dashcam dataset for roadside litter detection, capturing extremely small and sparse litter in dynamic driving contexts. The paper benchmarks accuracy-oriented transformers and real-time YOLO models to characterize localization precision versus inference speed, revealing that CO-DETR offers the strongest localisation (AP50:95) while real-time models lag on small-object detection. Key contributions include the dataset with 11,565 images and detailed statistical analyses, a comprehensive detector benchmark across 2021–2025 architectures, and insights into architectural trade-offs for edge deployability. The work provides a foundation for developing scalable, low-cost roadside-litter monitoring systems and highlights the need for high-resolution feature pathways to tackle extreme small-object regimes in real-world dashcam data.

Abstract

Roadside litter poses environmental, safety and economic challenges, yet current monitoring relies on labour-intensive surveys and public reporting, providing limited spatial coverage. Existing vision datasets for litter detection focus on street-level still images, aerial scenes or aquatic environments, and do not reflect the unique characteristics of dashcam footage, where litter appears extremely small, sparse and embedded in cluttered road-verge backgrounds. We introduce RoLID-11K, the first large-scale dataset for roadside litter detection from dashcams, comprising over 11k annotated images spanning diverse UK driving conditions and exhibiting pronounced long-tail and small-object distributions. We benchmark a broad spectrum of modern detectors, from accuracy-oriented transformer architectures to real-time YOLO models, and analyse their strengths and limitations on this challenging task. Our results show that while CO-DETR and related transformers achieve the best localisation accuracy, real-time models remain constrained by coarse feature hierarchies. RoLID-11K establishes a challenging benchmark for extreme small-object detection in dynamic driving scenes and aims to support the development of scalable, low-cost systems for roadside-litter monitoring. The dataset is available at https://github.com/xq141839/RoLID-11K.

Figures (8)

• Figure 1: Overview of the proposed RoLID-11K dataset. A vehicle-mounted dashcam serves as a mobile data acquisition platform, capturing roadside litter under diverse real-world driving conditions. The dataset comprises 11K annotated images spanning various weather, lighting, and road environments.
• Figure 2: Distribution of object counts per image for training, validation, and test splits in our RoLID-11K dataset. The distribution exhibits a long-tail pattern, reflecting real-world roadside litter scenarios, with most images containing 1–3 objects.
• Figure 3: Histogram of object areas in logarithmic scale across dataset splits. The peak around $\log_{10}(\text{Area}) \approx [2.4, 2.8]$ indicates that most litter objects occupy relatively small regions in the image, posing challenges for small object detection.
• Figure 4: Object size distribution following the COCO evaluation criteria: small, medium, and large. Small objects dominate across training (83.7%), validation (81.0%), and test (86.8%) splits, underscoring the challenges of small object detection in roadside litter scenarios.
• Figure 5: Scatter plot of bounding box dimensions (width vs. height) across dataset splits. The dashed line indicates a 1:1 aspect ratio. Training and validation sets exhibit concentrated distributions with similar patterns, while the test set shows more diverse shape variations and aspect ratios, providing a challenging benchmark for evaluating model robustness and generalization.