HSDA: High-Frequency Shuffle Data Augmentation for Bird's-Eye-View Map Segmentation

Calvin Glisson, Qiuxiao ChenWACV 2025

TL;DR

This paper introduces High-frequency Shuffle Data Augmentation (HSDA), a novel technique that leverages frequency domain manipulation to improve Bird's-Eye-View (BEV) map segmentation for autonomous driving. By strategically shuffling high-frequency image components, HSDA enhances a model's ability to discern fine details and edges, leading to state-of-the-art performance on the nuScenes dataset.

What Problem Does This Paper Address?

Autonomous driving systems rely heavily on accurate perception of their surroundings, and Bird's-Eye-View (BEV) map segmentation is a crucial component for this. BEV maps provide a top-down semantic understanding of the environment, classifying areas like drivable paths, pedestrian crossings, and stop lines. While data augmentation is a common practice to improve these segmentation models, existing methods primarily focus on manipulating the spatial domain of images. This leaves a gap in exploring how frequency domain information, which is vital for capturing sharp edges and intricate details, can be effectively leveraged for BEV map segmentation. The paper aims to address this by investigating the role of high-frequency components and proposing a novel augmentation strategy that exploits them.

Methods

The core of this paper is the High-frequency Shuffle Data Augmentation (HSDA) method. HSDA operates on individual camera images by:

  1. Frequency Decomposition: An image is converted to the frequency domain using the Fast Fourier Transform (FFT). The resulting spectrum is then decomposed into low-frequency and high-frequency components using Gaussian low-pass and high-pass filters.
  2. High-Frequency Shuffling: The dominant high-frequency components are identified based on their magnitude. A subset of these top K frequencies are then randomly shuffled within the high-frequency spectrum. This introduces controlled noise that perturbs fine details without destroying the overall structure.
  3. Recombination and Spatialization: The shuffled high-frequency spectrum is recombined with the original low-frequency spectrum. The inverse FFT (iFFT) is then applied to convert this back into the spatial domain, resulting in an augmented image.
  4. Integration: This augmented image replaces the original image during training. The key insight is that this process encourages the network to learn the correlation between high-frequency elements and the BEV map, improving its ability to distinguish important details from noise.

HSDA is designed to be model-agnostic and requires no modifications to the baseline network architecture or additional parameters.

Datasets

The primary dataset used for training and evaluation is nuScenes, a large-scale autonomous driving dataset containing 1,000 scenes with multi-view camera images and semantic map annotations. The authors also mention using the KITTI dataset for evaluating the applicability of HSDA to monocular 3D object detection.

Key Insights and Ideas

  • High-frequency information is critical for BEV segmentation: The paper empirically demonstrates that models relying solely on low-frequency components suffer significant accuracy loss, while those using high-frequency components maintain higher performance. This highlights the importance of edges, textures, and fine details for accurately delineating regions in BEV maps.
  • Frequency domain augmentation is underexplored: While spatial domain augmentations are common, the potential of frequency domain manipulation for BEV segmentation was largely untapped. HSDA bridges this gap by introducing a novel way to leverage frequency information.
  • Controlled noise injection improves robustness: By shuffling high-frequency components, HSDA introduces a form of controlled noise that forces the model to learn more robust representations, distinguishing essential high-frequency features from spurious ones.
  • HSDA is broadly applicable: The method shows consistent performance improvements across different BEV segmentation architectures (BEVFusion, Baseline, RGC) and even extends to a different task (monocular 3D object detection).

Key Results

  • State-of-the-art performance: When applied to the RGC model on the nuScenes dataset, HSDA achieved a new state-of-the-art mean Intersection over Union (mIoU) of 61.3% for camera-only systems, surpassing previous approaches by at least 1.6%.
  • Consistent improvements: HSDA consistently improved mIoU by 1.6% for the Baseline model and 1.6% for the RGC model compared to their non-augmented counterparts.
  • Broad class improvement: The augmentation improved performance across nearly all semantic classes, with no instances of accuracy degradation.
  • Effective parameter tuning: Ablation studies identified optimal values for the key parameters K (number of shuffled frequencies) and D (filter cutoff), with K=2000 and D=10 proving most effective.
  • Monocular 3D Object Detection gains: HSDA also yielded substantial improvements in pedestrian and car detection when applied to the MonoCon model on the KITTI dataset, demonstrating its versatility.

Limitations and Future Work

  • Nighttime performance: While HSDA shows promising results in daytime and rainy conditions, the paper acknowledges that nighttime performance exhibits greater uncertainty, particularly at far distances. This is attributed to reduced light and potential camera anomalies.
  • Further exploration of frequency domain: The authors suggest that their work opens avenues for further exploration of frequency domain applications in autonomous driving, implying that other frequency-based augmentation or processing techniques could be beneficial.
  • Extension to other tasks: The successful application to monocular 3D object detection suggests HSDA could be valuable for other perception tasks that rely on fine-grained visual details.

Summary updated

BibTeX
@inproceedings{glisson2025hsda,
  author = {Calvin Glisson and Qiuxiao Chen},
  title = {{HSDA: High-Frequency Shuffle Data Augmentation for Bird's-Eye-View Map Segmentation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure