AI News AI资讯 2d ago Updated 2d ago 更新于 2天前 52

Ant Group’s Robbyant Open-Sources LingBot-Vision: A 1B Boundary-Centric Vision Foundation Model for Dense Spatial Perception 蚂蚁集团Robbyant开源LingBot-Vision:一款面向密集空间感知的10亿参数边界中心视觉基础模型

Ant Group’s Robbyant open-sourced LingBot-Vision, a 1B-parameter self-supervised Vision Transformer optimized for dense spatial perception via boundary-centric pretraining. The model introduces "Masked Boundary Modeling," forcing the network to prioritize object boundaries during training, enabling it to outperform significantly larger models like the 7B DINOv3 on depth estimation and segmentation tasks. Training utilizes a novel categorical boundary field and an automatic Number-of-False-Alarms Ant Group旗下Robbyant开源LingBot-Vision,这是一个专为密集空间感知设计的自监督视觉Transformer家族,包含ViT-giant至ViT-small四种规模。 核心创新在于“掩码边界建模”(Masked Boundary Modeling),将物体边界作为原生预训练信号而非下游任务,使1B参数模型在密集空间任务上超越甚至匹敌7倍大小的DINOv3。 采用边界强制策略和分类边界场技术,结合NFA无参数假阳性检验,实现了语义与几何表示的协同涌现,且训练数据量仅为DINOv3的三分之一。 在NYU-Depth v2、KITTI等基准测试中表现优异,无需训练即可通过标

72
Hot 热度
78
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • Ant Group’s Robbyant open-sourced LingBot-Vision, a 1B-parameter self-supervised Vision Transformer optimized for dense spatial perception via boundary-centric pretraining.
  • The model introduces "Masked Boundary Modeling," forcing the network to prioritize object boundaries during training, enabling it to outperform significantly larger models like the 7B DINOv3 on depth estimation and segmentation tasks.
  • Training utilizes a novel categorical boundary field and an automatic Number-of-False-Alarms (NFA) validation mechanism, allowing the model to achieve state-of-the-art results on a fraction of the data required by competitors.
  • The architecture includes distilled variants (ViT-L, ViT-B, ViT-S) and demonstrates exceptional performance in video object segmentation without temporal supervision, leveraging stable boundary token tracking.

Why It Matters

This release challenges the prevailing paradigm in vision foundation models that prioritizes semantic invariance over spatial structure, offering a highly efficient alternative for applications requiring precise geometric understanding. By demonstrating that a 1B-parameter model can match or exceed 7B-parameter counterparts, it provides a cost-effective solution for resource-constrained environments such as robotics and autonomous driving. The open-source availability under Apache-2.0 accelerates research into dense spatial perception and embodied AI.

Technical Details

  • Architecture and Scale: The flagship model is a ViT-g/16 with approximately 1.1B parameters, trained on a curated corpus of 161M images. Smaller variants (ViT-L: 300M, ViT-B: 86M, ViT-S) are derived via distillation.
  • Masked Boundary Modeling: Unlike standard random masking, this method identifies boundary-bearing tokens using a teacher network and forces them into the masked set. These tokens receive an explicit geometric target alongside the semantic self-distillation target, resolving ambiguity at object edges.
  • Categorical Boundary Field: Boundaries are represented as line segments with attribute vectors (distance and angles), discretized into 32 bins per channel for classification. This allows the use of centering/sharpening techniques and enables a parameter-free NFA test to validate predicted structures, ensuring only supported geometry serves as a teaching signal.
  • Training Objective: The loss function combines standard DINO ($L_{DINO}$), iBOT ($L_{iBOT}$), boundary loss ($L_{bnd}$), and KoLeo regularization ($L_{KoLeo}$), trained without human labels or external edge detectors.

Industry Insight

  • Efficiency in Spatial Tasks: Developers working on robotics or AR/VR should consider LingBot-Vision as a drop-in replacement for larger semantic models when dense spatial accuracy (depth, segmentation) is critical, potentially reducing compute costs by up to 7x.
  • Shift in Pretraining Strategies: The success of boundary-centric pretraining suggests that future foundation models may increasingly incorporate geometric priors directly into self-supervised objectives rather than relying solely on semantic reconstruction.
  • Low-Resource Deployment: The availability of distilled versions and the ability to perform training-free video object segmentation make this model particularly suitable for edge devices and real-time applications where heavy temporal supervision is impractical.

TL;DR

  • Ant Group旗下Robbyant开源LingBot-Vision,这是一个专为密集空间感知设计的自监督视觉Transformer家族,包含ViT-giant至ViT-small四种规模。
  • 核心创新在于“掩码边界建模”(Masked Boundary Modeling),将物体边界作为原生预训练信号而非下游任务,使1B参数模型在密集空间任务上超越甚至匹敌7倍大小的DINOv3。
  • 采用边界强制策略和分类边界场技术,结合NFA无参数假阳性检验,实现了语义与几何表示的协同涌现,且训练数据量仅为DINOv3的三分之一。
  • 在NYU-Depth v2、KITTI等基准测试中表现优异,无需训练即可通过标签传播在视频对象分割任务中达到与大型模型相当的性能。

为什么值得看

这篇文章展示了如何通过改变预训练目标(从语义不变性转向空间结构感知)来显著提升视觉基础模型在机器人和具身智能关键任务上的效率。对于AI从业者和行业而言,它证明了小参数模型可以通过更精细的空间表征学习,以极低的计算成本实现媲美超大模型的密集预测能力,为资源受限场景下的部署提供了新范式。

技术解析

  • 架构与规模:旗舰模型ViT-g/16拥有约11亿参数,并蒸馏出ViT-L(3亿)、ViT-B(8600万)和ViT-S版本。所有模型均输出用于冻结读取器的密集patch-token特征,权重以Apache-2.0协议开源。
  • 掩码边界建模机制:区别于标准随机掩码,该方法通过教师网络在线预测密集边界场,识别出包含边界的token并将其强制加入学生网络的掩码集合。边界token接收显式几何目标加语义自蒸馏目标,而内部区域仅保留语义目标,解决了两类表示的竞争问题。
  • 分类边界场与NFA验证:将边界建模为包含距离和角度的属性向量,并通过离散化为32个bin转化为像素级分类问题。利用“无结构”假设下方向均匀分布的特性,引入无参数的假阳性警报(NFA)测试验证解码出的线段,确保只有被支持的真实结构成为教学信号。
  • 训练效率与数据:使用约1.61亿张精选图像进行训练,无人工标签、无外部边缘检测器。训练样本量比DINOv3少一个数量级,显著降低了算力需求。
  • 基准性能:在NYU-Depth v2上RMSE达0.296(优于7B DINOv3的0.309);在ADE20K、Cityscapes等语义分割任务上,1B模型性能接近或超过更大规模的对比模型;在视频对象分割中,仅凭冻结特征的余弦相似度跟踪即可达到70.0 J&F(DAVIS-2017)。

行业启示

  • 空间感知是具身智能的关键瓶颈:当前主流视觉模型过度关注语义分类,忽视了边界和深度等空间结构信息。LingBot-Vision的成功表明,将空间结构作为预训练的核心信号,能极大提升模型在机器人导航、操作等物理交互任务中的实用性。
  • 高效预训练策略的潜力:通过创新的损失函数设计(如边界强制和NFA验证),可以用更少的数据和算力训练出性能更强的模型。这为降低大模型训练成本、加速迭代提供了新的技术路径。
  • 开源生态对技术扩散的推动:Ant Group开源该模型及其代码,有助于学术界和工业界快速验证基于边界感知的视觉表征方法,推动密集预测任务在更多垂直领域的落地应用。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Multimodal 多模态 Robotics 机器人