HeAL3D: Heuristical-enhanced Active Learning for 3D Object Detection (2505.00507v2)

Abstract: Active Learning has proved to be a relevant approach to perform sample selection for training models for Autonomous Driving. Particularly, previous works on active learning for 3D object detection have shown that selection of samples in uncontrolled scenarios is challenging. Furthermore, current approaches focus exclusively on the theoretical aspects of the sample selection problem but neglect the practical insights that can be obtained from the extensive literature and application of 3D detection models. In this paper, we introduce HeAL (Heuristical-enhanced Active Learning for 3D Object Detection) which integrates those heuristical features together with Localization and Classification to deliver the most contributing samples to the model's training. In contrast to previous works, our approach integrates heuristical features such as object distance and point-quantity to estimate the uncertainty, which enhance the usefulness of selected samples to train detection models. Our quantitative evaluation on KITTI shows that HeAL presents competitive mAP with respect to the State-of-the-Art, and achieves the same mAP as the full-supervised baseline with only 24% of the samples.

HeAL3D: Heuristical-enhanced Active Learning for 3D Object Detection

The paper presents a novel approach called HeAL (Heuristical-enhanced Active Learning) for active learning specifically applied to 3D object detection using LiDAR data. The core motivation behind this work is to address the challenge of data labeling in uncontrolled scenarios for autonomous driving, where the labeling process is not only time-consuming but also resource-intensive. Unlike previous approaches that concentrate largely on theoretical problem-solving methodologies for sample selection, this paper emphasizes practical insights by integrating heuristical features that leverage the extensive literature on 3D detection models.

HeAL differentiates itself by supplementing traditional uncertainty metrics with heuristical features to determine sample informativeness more effectively. The proposed method considers object distance and point-quantity—two factors that are directly retrievable from LiDAR data and significantly affect detection uncertainty due to their impact on object localization and classification. This consideration is crucial because objects located further from the sensor or those consisting of fewer points—characteristics inherent to LiDAR data—typically present greater detection challenges.

A key component of HeAL is its application of a Gaussian Mixture Model (GMM) that represents the scene's probability distribution. This representation allows for an agnostic approach to model architecture, enabling more flexible implementation across different 3D detection systems. The GMM framework is employed to estimate inconsistency scores between original and augmented samples, leveraging both class and distance features to augment existing uncertainty estimates. Thus, HeAL strategically selects samples that offer the most significant contribution to the learning process while requiring reduced labeling efforts.

Quantitatively, HeAL demonstrates its efficacy using the KITTI benchmark dataset, achieving competitive mean Average Precision (mAP) compared to state-of-the-art methods. Notably, it reaches the same mAP as a fully supervised baseline model using only 24% of the labeled samples. These results underscore the practical benefits of incorporating heuristical knowledge directly into the active learning paradigm for 3D object detection and highlight the potential for substantial labeling cost reductions without compromising model performance.

The introduction of HeAL signifies an intriguing shift towards integrating domain-specific heuristics for active learning, particularly in scenarios where data labeling is impractical or expansive. This work not only advances the practices in active learning strategies but also stimulates further exploration into how heuristic domain knowledge can be systematically integrated into machine learning frameworks to optimize performance in real-world applications. Future research directions may include expanding HeAL's applicability to other LiDAR datasets and further exploring the balance between heuristic insights and model-agnostic techniques to continue improving efficiency in data-driven tasks.