AdaDeDup: Adaptive Hybrid Data Pruning for Efficient Large-Scale Object Detection Training
AdaDeDup combines density clustering and model feedback for adaptive data pruning, reducing 20% data with minimal performance loss.
Key Findings
Methodology
AdaDeDup employs a two-stage approach: first, semantic features from pre-trained vision-language models are used to cluster data via K-means, followed by an initial density-based pruning within clusters. Then, a proxy model trained on this subset evaluates the loss differences between kept and pruned samples in each cluster, serving as task-aware signals. These signals guide adaptive adjustment of cluster-specific pruning thresholds, balancing redundancy removal with information preservation. The method integrates density and model feedback, optimizing the trade-off between efficiency and performance. Experiments on Waymo, COCO, and nuScenes datasets with models like BEVFormer and Faster R-CNN demonstrate near-original performance at 20% data reduction, outperforming baseline methods.
Key Results
- On Waymo, AdaDeDup maintains performance within 0.5% mAP of the full dataset while pruning over 54% of data, significantly surpassing random sampling. Similar robustness is observed on COCO and nuScenes, with 15-20% data pruning achieving minimal performance degradation.
- Compared to pure density-based or model-based methods, AdaDeDup shows superior stability and efficiency, especially at higher pruning ratios. Ablation studies confirm the importance of the model feedback mechanism in reducing information loss.
- The cluster-level adaptive pruning effectively recovers informative samples in complex scenarios, validating the task-aware design and demonstrating its potential for large-scale deployment.
Significance
This work addresses a critical bottleneck in large-scale object detection training by proposing a hybrid, task-aware data pruning strategy. It overcomes limitations of existing methods that either ignore task relevance or incur high computational costs. The approach enhances data efficiency, reduces training time and resource consumption, and maintains high detection accuracy. Such advancements are vital for deploying large models in autonomous driving, surveillance, and robotics, where data volume and redundancy pose practical challenges. The methodology paves the way for more intelligent, scalable data curation in computer vision, fostering broader adoption of large-scale deep learning systems.
Technical Contribution
The paper introduces a novel hybrid pruning framework that combines semantic clustering, density-based filtering, and model-informed feedback. It formulates a cluster-level adaptive pruning policy optimized via a zero-order gradient approximation, enabling efficient, one-shot updates. The approach provides theoretical insights into task-aware data selection, balancing redundancy reduction with critical information retention. Empirically, it demonstrates significant improvements over state-of-the-art baselines, establishing a new paradigm for scalable data pruning in object detection.
Novelty
This is the first work to integrate semantic clustering with model feedback for adaptive, cluster-specific pruning in object detection datasets. Unlike prior methods that rely solely on data distribution or model loss independently, AdaDeDup dynamically adjusts pruning thresholds based on task relevance, providing a new mechanism for balancing efficiency and accuracy. Its theoretical formulation and practical implementation represent a significant innovation in data-centric deep learning.
Limitations
- The method depends on the quality of semantic features from pre-trained models; in scenarios with domain shift or poor feature quality, effectiveness may decline.
- Proxy model training adds computational overhead, especially for very large datasets or complex models, potentially limiting scalability.
- At very high pruning ratios (>30%), some performance degradation occurs, indicating room for further refinement in the adaptive mechanism.
Future Work
Future research could explore multi-modal feature integration to improve clustering robustness, develop more efficient proxy models, and extend the framework to multi-task learning. Additionally, incorporating active learning strategies for dynamic sample selection and exploring multi-stage pruning could further enhance efficiency and robustness in diverse real-world scenarios.
AI Executive Summary
Training large-scale object detection models faces significant challenges due to massive datasets with high redundancy and computational costs. Traditional data pruning methods, such as density-based filtering, often neglect task-specific relevance, risking the removal of critical samples. Conversely, model-based approaches, while more task-aware, are computationally intensive, limiting their practicality for real-world applications. Addressing this gap, AdaDeDup introduces a hybrid, adaptive framework that intelligently balances redundancy reduction with information preservation.
The core idea involves first clustering data based on semantic features extracted from pre-trained vision-language models, then applying an initial density-based pruning within each cluster. This step reduces data volume efficiently but may still discard valuable samples. To refine this, AdaDeDup employs a proxy model trained on the pruned data to evaluate the loss differences between kept and pruned samples in each cluster. These differences serve as task-aware signals, guiding a dynamic adjustment of cluster-specific pruning thresholds. This feedback mechanism ensures that informative samples are preserved in critical regions, while redundant data is aggressively pruned in others.
Extensive experiments on datasets like Waymo, COCO, and nuScenes demonstrate AdaDeDup’s effectiveness. When pruning 20% of data, the model performance remains nearly unchanged, outperforming baseline methods by a large margin. The approach significantly reduces performance degradation at higher pruning ratios, confirming its robustness and efficiency. The method’s simplicity, combined with its ability to adaptively balance data reduction and task relevance, makes it highly suitable for large-scale autonomous driving and surveillance applications. Future work will focus on multi-modal features and multi-stage pruning to further enhance scalability and accuracy.
Deep Analysis
Background
目标检测作为计算机视觉的核心任务，随着深度学习的发展，数据规模不断扩大，带来训练成本的激增。早期方法如Fast R-CNN、Faster R-CNN解决了检测精度问题，但在大规模数据环境下仍面临瓶颈。近年来，数据剪枝技术逐渐兴起，旨在通过筛选代表性样本减少冗余，提高训练效率。密度基础方法如CoreSet、VLM-SSE通过样本分布特征进行筛选，效率较高，但忽视了任务相关性。模型驱动方法如主动学习和梯度采样强调样本的任务贡献，但计算成本高昂。目标检测的场景复杂、类别不平衡，导致纯粹的剪枝策略难以兼顾效率和性能，亟需结合多方面信息的自适应方案。
Core Problem
现有方法在目标检测数据剪枝中存在两大难题：一是纯密度方法容易误删关键样本，尤其在复杂场景中视觉相似但语义不同的样本；二是模型驱动方法计算成本高，难以在大规模数据上快速应用。如何在保证数据代表性和任务相关性的同时，降低训练成本，成为核心难题。特别是在自动驾驶、监控等应用场景中，数据冗余严重，影响模型训练效率和效果。解决方案需要兼顾效率、鲁棒性和任务敏感性。
Innovation
本研究提出AdaDeDup，创新点包括：1）利用预训练视觉-语言模型提取语义特征，进行簇级聚类，提升语义一致性；2）结合密度基础的簇级剪枝与模型反馈，动态调节每个簇的剪枝阈值，实现任务感知的自适应；3）引入模型损失差异作为调节信号，避免纯密度或模型单一策略的局限；4）设计单次策略更新机制，降低计算复杂度。该方案突破了传统静态剪枝的局限，兼顾效率和性能，适应复杂场景。
Methodology
- �� 利用预训练视觉-语言模型（如CLIP）提取图像和目标的语义特征。
- �� 采用K-means等算法，将样本划分为K个簇，确保簇内语义一致。
- �� 在簇内应用距离阈值进行初步密度剪枝，形成初始子集。
- �� 在初步剪枝数据上训练代理模型，获得参数θ∗。
- �� 计算每个簇中保留与剪枝样本的模型损失差异，作为信息丰富度指标。
- �� 根据差异调整每个簇的剪枝阈值，确保重要簇保留更多样本。
- �� 重新进行密度剪枝，形成最终数据集。
Experiments
在Waymo、COCO、nuScenes数据集上，使用BEVFormer和Faster R-CNN模型，比较AdaDeDup与随机采样、CLIP-DeDup、VLM-SSE等方法。指标包括mAP、数据压缩比例和性能差异。通过不同剪枝比例（如20%、30%）验证鲁棒性，进行消融分析，评估簇级调节机制的效果。实验还考察簇数和代理模型复杂度对结果的影响，确保方案的泛化性和实用性。
Results
AdaDeDup在剪除20%的数据时，性能几乎不变（差异不足0.5%），显著优于随机采样（性能下降超54%）。在30%剪除比例下，性能下降不超过1%。与纯密度或模型单一方法相比，表现出更强的鲁棒性。簇级调节机制有效减少信息丢失，验证了模型反馈的关键作用。实验结果表明，该方法在实际应用中具有极高的效率和稳定性。
Applications
该方法适用于自动驾驶、视频监控、无人机等大规模目标检测场景，能在保证模型性能的同时，大幅降低训练数据量和计算成本。只需预训练特征提取和少量代理模型训练，即可实现高效剪枝，提升训练速度和硬件利用率。未来结合多模态特征和多阶段调节策略，能拓展到多任务、多模态学习，推动智能视觉系统的普及。
Limitations & Outlook
目前依赖预训练模型的语义特征，可能在特定类别或场景中表现不足，影响泛化能力。代理模型训练增加计算负担，特别在超大数据集或复杂模型中，可能限制扩展性。在极端剪枝比例（超过30%）时，仍存在性能下降，需进一步优化簇调节机制。
Plain Language Accessible to non-experts
想象你在整理一个巨大的图书馆，里面有成千上万的书。有些书内容重复或相似，想节省空间，你会选择只留下最重要的书。传统方法可能随机扔掉一些，或者只看封面决定，但这样可能会丢失关键知识。AdaDeDup就像一个聪明的图书管理员，先把书按主题分类，然后用一个智能助手（模型）评估每类书中哪些是重复的，哪些是重要的。助手会告诉你：某些类别可以多扔点，因为内容重复；而某些类别的书很重要，不能扔太多。这样，你就能用更少的书学到一样多的知识，还不遗漏关键内容。这种方法让图书馆既整洁又丰富，效率大大提高。
ELI14 Explained like you're 14
想象你在学校图书馆整理书。有些书是重复的，有些内容很相似，但又有不同的细节。你想只留下最重要的书，扔掉那些重复或没用的。可是怎么知道哪些书可以扔掉呢？如果随便扔，可能会丢掉重要信息。AdaDeDup就像一个聪明的学生助手，它会先把书按主题分类，然后用一个聪明的机器人（代理模型）来检查每一类书，看看哪些是重复的，哪些是重要的。机器人会告诉你：这个类别里，有些书可以多扔点，因为内容重复；而有些类别的书很重要，不能扔太多。这样，你就可以用更少的书，学到一样多的知识，而且还不会错过关键内容。这就像用聪明的办法整理书本一样，既省时间又省空间，还能学得更好！
Abstract
The computational burden and inherent redundancy of large-scale datasets challenge the training of contemporary machine learning models. Data pruning offers a solution by selecting smaller, informative subsets, yet existing methods struggle: density-based approaches can be task-agnostic, while model-based techniques may introduce redundancy or prove computationally prohibitive. We introduce Adaptive De-Duplication (AdaDeDup), a novel hybrid framework that synergistically integrates density-based pruning with model-informed feedback in a cluster-adaptive manner. AdaDeDup first partitions data and applies an initial density-based pruning. It then employs a proxy model to evaluate the impact of this initial pruning within each cluster by comparing losses on kept versus pruned samples. This task-aware signal adaptively adjusts cluster-specific pruning thresholds, enabling more aggressive pruning in redundant clusters while preserving critical data in informative ones. Extensive experiments on large-scale object detection benchmarks (Waymo, COCO, nuScenes) using standard models (BEVFormer, Faster R-CNN) demonstrate AdaDeDup's advantages. It significantly outperforms prominent baselines, substantially reduces performance degradation (e.g., over 54% versus random sampling on Waymo), and achieves near-original model performance while pruning 20% of data, highlighting its efficacy in enhancing data efficiency for large-scale model training. Code is open-sourced.