# RangeSAM：视觉基础模型在距离视图激光雷达分割中的潜力 Paul Julius Kühn Duc Anh Nguyen Arjan Kuijper Saptarshi Neil Sinha ## 摘要 激光雷达点云分割是自动驾驶和3D场景理解领域的核心任务。尽管基于体素和点的方法因能与深度架构兼容且能捕捉精细的几何特征，在近期研究中占据主导地位，但它们往往存在计算成本高、内存访问不规律以及因规模扩展问题导致的运行时效率受限等问题。相比之下，距离视图方法虽然研究相对较少，却能够借助成熟的2D语义分割技术实现快速准确的预测。受视觉基础模型（VFMs）在图像描述、零样本识别及多模态任务中取得的快速进展启发，我们探究了当前用于分割任务的顶尖视觉基础模型SAM2，是否可作为距离视图下激光雷达点云分割的强大骨干网络。据我们所知，RangeSAM是首个将SAM2适配于3D分割的距离视图框架，它结合了高效的2D特征提取技术与投影/反投影操作来处理点云数据。为了优化SAM2在距离视图表示下的性能，我们对编码器做了多项架构调整：（1）设计了一种全新的Stem模块，用以强化激光雷达距离图像中固有的水平空间依赖关系；（2）针对球面投影的几何特性定制了Hiera Blocks的配置；（3）在编码器骨干网络中采用了适配后的窗口注意力机制，专门用来捕捉距离视图伪图像中特有的空间模式与不连续性。我们的方法在SemanticKITTI数据集上取得了具有竞争力的性能，同时还受益于以2D为中心的处理流程所具备的速度、可扩展性及部署简便性。本研究证明了视觉基础模型作为点云分割通用骨干网络的可行性，为基于基础模型的统一激光雷达分割研究开辟了新路径。研究结果让我们得出结论：采用视觉基础模型的距离视图分割方法能够取得令人鼓舞的成果。 ### 1. 引言 利用二维和三维语义信息重建城市场景对于自动驾驶系统（车辆、无人机、机器人）而言至关重要，这要求实现实时视点合成、语义分割、深度生成以及动态物体跟踪。其中，激光雷达点云语义分割是区分车辆、行人、路标和基础设施的基础。近期的研究主要集中于基于体素和点的方法[^26] [^43] [^63] [^76] [^77] [^95]，这些方法虽然性能优异，但在处理大规模户外数据时会带来巨大的计算和内存开销[^18] [^46] [^47] [^49] [^64] [^67] [^70] [^75] [^85]，并且难以应对不规则、无序的点云数据。相比之下，距离视图分割[^2] [^13] [^33] [^41] [^48]将三维点云投影为密集的二维表示，从而能够复用成熟的二维模型，进而降低内存占用并加快推理速度。尽管由于在处理遮挡现象和分辨率损失方面存在局限，距离视图分割此前未被重视，但注意力机制[^66]、多尺度融合[^93]以及上下文感知架构[^40]的最新进展使得我们有理由重新评估这种距离视图范式。我们提出了RangeSAM，它将Segment Anything Model 2（SAM2）[^50]应用于距离视图激光雷达分割任务中（图1）。我们的方法利用了SAM2在二维理解方面的强大零样本能力，并将其通过距离视图表示扩展到三维领域。该方法的具体步骤包括：（1）进行距离投影预处理，以转换无序的激光雷达扫描数据；（2）采用包含感受野块的多组件架构[^37]；（3）利用k-NN标签传播进行后处理；（4）设计复合损失函数以解决类别不平衡和边界精度问题。我们的主要贡献如下： - 我们提出了RangeSAM，这是首个通过距离视图表示将SAM2应用于激光雷达点云分割的框架。 • 我们设计了多组件编码器架构，该架构采用了预训练的Hiera骨干网络、定制的Stem模块、新颖的嵌入矩阵，以及利用局部和全局多头注意力的Hiera块（图1）。 --- • 我们在SemanticKITTI数据集上验证了该方法具有具竞争力的性能[^4]，同时针对训练策略和数据增强方式开展了消融研究。 ### 2. 相关工作 ## 视觉基础模型 近期，大规模视觉模型彻底革新了图像任务 [^6] [^38] [^79] [^88]。Segment Anything Model（SAM）系列实现了具备零样本迁移能力的可提示分割 [^30] [^36] [^53] [^56]；而SAM2则进一步扩展至视频和复杂场景 [^12] [^28] [^39] [^50] [^80] [^87]，并且已被应用于三维领域 [^7] [^83] [^86]。其他视觉基础模型还包括自监督Transformer（DINO系列 [^9] [^42] [^57]）、多模态模型 [^45] [^97] 以及EVA系列 [^20] [^21] [^60] [^61]，它们具备强大的泛化能力和标签高效学习能力。例如，DINOv2的特征可提升三维语义分割效果 [^83] [^89]；而CLIP的图像-文本嵌入则能够实现零样本三维理解 [^11] [^45] [^62] [^90]。 ### 2.1. 基于投影表示的方法 卷积神经网络（CNN）：基于投影的方法将点云转换为结构化的二维表示（距离视图、鸟瞰图或透视视图），从而利用成熟的二维CNN架构 [^4] [^5] [^10] [^16] [^27] [^31] [^34] [^41] [^51] [^64] [^65] [^81] [^91] [^96]。SqueezeSeg系列 [^68] [^72] [^73] [^81]率先采用轻量级CNN与CRF后处理技术实现实时分割 [^17]。DarkNetSeg [^4]、SalsaNext [^16]、KPRNet [^31]、RangeNet++ [^41]、Lite-HDSeg [^51] 和 SqueezeSegV3 [^81]等方法均采用两阶段方案：先进行距离投影，再采用U-Net风格的架构 [^54]。RangeSeg [^13] [^69]则针对球面投影引入了多尺度上下文建模机制。 视觉Transformer（ViT）：ViT架构推动了重大进展：RangeViT和FRNet [^2] [^84]提出了结合卷积归纳偏置与全局注意力的混合CNN-Transformer骨干网络，从而受益于二维领域的迁移学习 [^15]；而RangeFormer [^33]则构建了纯Transformer架构，在室外基准测试中取得了顶尖性能 [^4] [^5] [^8] [^44] [^59]。 ### 3. 方法论 我们提出了RangeSAM——一种基于投影的激光雷达分割模型，该模型充分利用了视觉基础模型的优势。我们详细阐述了距离投影预处理、架构修改以及后处理流程。图1展示了整个流程，包括词元模块、基于SAM2的编码器和解码器组件。 #### 3.1. 距离投影 每次激光雷达扫描最初都表示为一个无序的点集 $\mathcal{P} = (x, y, z, f)$，其中 $(x, y, z)$ 表示笛卡尔坐标，f 代表辅助传感器测量值，如强度或反射值。随后，点云通过投影到传感器的球面坐标系 [^41] 来转换为距离视图表示： $$ \theta=\arctan\left(\frac{y}{x}\right),\phi=\arcsin\left(\frac{z}{r}\right),r=\sqrt{x^{2}+y^{2}+z^{2}} $$ 具体来说，3D点通过光栅化被离散化为尺寸为 $H \times W$ 的2D圆柱投影： $$ \binom{u}{v}=\binom{\frac{1}{2}\left[1-\arctan(y,x)\pi^{-1}\right]w}{\left[1-\left(\arcsin\left(zr^{-1}\right)+\mathfrak{f}_{\mathrm{u p}}\right)\mathfrak{f}^{-1}\right]h}, $$ 其中 $(u,v)$ 表示点 $p_n$ 在距离图像坐标系中的投影坐标，h 和 w 分别表示图像的高度和宽度参数。$r=\sqrt{x^2+y^2+z^2}$ 表示点与传感器之间的距离，传感器的垂直视场角 f 由 $(f_{\mathrm{up}} + f_{\mathrm{down}})$ 个视角之和构成。在我们的方法中，我们将成像系统配置为以常用的空间分辨率 $64\times2048$ 像素运行 [^2] [^13] [^33]。对于投影到同一像素的多个点，我们通过保留最小距离特征来解决冲突；未投影的像素则填充为零。 #### 3.2. 模型架构 SAM2是一种在SA-V数据集上训练的最先进的视觉基础模型 [^50]。SAM2-UNet [^80] 在保持简洁性的同时，在多项分割任务中表现出卓越的性能。RangeSAM采用了类似的范式（图1），引入了感受野块（RFB）[^37] 以增强特征解码能力。主干网络：通过线性变换、层归一化 [^3] 和GELU激活函数 [^25] 将输入张量从 (B,6,H,W) 转换为 (B,96,H,W)（图1顶部）；随后，特征被划分为步长为1的重叠7×7块。我们用一种新颖的 (4,128) 嵌入矩阵替换了SAM2的位置嵌入，以增强水平空间敏感性。编码器：采用预训练的Hiera [^50] [^55] 骨干网络，其多组件架构（图1右侧）包括：Hiera块：骨干网络的每个阶段包含指定数量的Hiera块，每个块包含两个模块：1) 多头注意力模块——在早期阶段，注意力被限制在窗口内，而在后期阶段也可能使用全局注意力。该操作可表示为 $$ X_{out}=X+\mathrm{DropPath}(\mathrm{MHA}(\mathrm{LayerNorm}(X))) $$ $$ \operatorname{MHA}(Q,K,V)=\operatorname{Concat}(\operatorname{head}_{1},\ldots,\operatorname{head}_{h})W^{O} $$ --- ![Image](images/c1f5f33df83e65377c86f9d1f82d1bfacdede4960542ecc14bc8619767c0d231.jpg)
Figure 1. Overview of our SAM2-based point-cloud segmentation model. The stem reshapes range-view point clouds into a tensor suitable for the encoder. The encoder is built from stacked Hiera blocks—each containing a multi-head self-attention module and a feed-forward network [^50]. The decoder comprises of Receptive Field Blocks [^37] (RFB) with LayerNorm [^3] and GELU [^25], concatenates multi-scale features, and projects them to $N_{classes}$ while also adding auxiliary head (Aux) on corresponding output.
其中 $$ \mathrm{h e a d}_{i}=\mathrm{A t t e n t i o n}\left(Q W_{Q}^{i},K W_{K}^{i},V W_{V}^{i}\right), $$ $$ Attention(Q,K,V)=\sigma\left(\frac{QK^{\top}+M}{\sqrt{d_{head}}}\right)V $$ 其中DropPath代表随机深度正则化技术 [^35]，LayerNorm表示层归一化方法 [^3]，它提供了与批量大小无关的归一化能力。2) 前馈网络： $$ O=\operatorname{D r o p P a t h}(\operatorname{M L P}(\operatorname{L a y e r N o r m}(O)))\oplus X_{\operatorname{o u t}}\,, $$ $$ F=\mathrm{Lin}\left(\mathrm{GELU}\left(\mathrm{DWConv2D}_{3\times3}(\mathrm{GELU}(\mathrm{O}))\right)\right)+\mathrm{O} $$ 其中DWConv2D $_{3\times3}$ 表示 $3\times3$ 深度卷积操作，它在保持极低参数开销的同时引入了空间局部性归纳偏置，从而确保计算效率。SAM2-tiny骨干架构包含四个分层阶段，分别含有 [^1] [^2] [^7] [^2] 个Hiera块。在块 [^5] [^7] [^9] [^11] 处实现了全局注意力机制，以捕捉长距离空间依赖关系。 注意力窗口：在早期阶段，Hiera采用局部窗口注意力机制，并配合掩码策略：当词元i和j属于同一掩码单元时，权重为$M_{ij} = 0$；否则权重为$M_{ij} = -\infty$。后期阶段则采用全局注意力机制，权重为$M = 0$。考虑到距离视图图像的分辨率为$64 \times 2048$，我们提出了一种强调水平空间关系的非对称注意力窗口设计：在第一和第四阶段使用$8 \times 64$，在第二和第三阶段使用$16 \times 128$。这种横向拉长的窗口配置虽源于经验性考量，但相比传统的方形注意力窗口展现出更优的性能，能够有效捕捉距离投影激光雷达数据固有的水平结构。 解码器：编码器的输出为多尺度特征$F_{1}, F_{2}, F_{3}, F_{4}$，其空间尺寸为… $$ \left[(H,W),\left(\frac{H}{2},\frac{W}{2}\right),\left(\frac{H}{4},\frac{W}{4}\right),\left(\frac{H}{8},\frac{W}{8}\right)\right] $$ 通道维度为[^96] [^192] [^384] [^768]。遵循[^80]的做法，我们使用经过改进的感受野模块[^37]架构，将输出通道维度统一设为256；该架构用LayerNorm和GELU激活函数替代了传统的BatchNorm[^29]与ReLU，从而更好地适配当代Transformer架构。按照[^22]的思路，我们将这四组归一化后的特征图拼接起来，逐步降低维度以匹配目标类别数量，同时在对应的特征层加入辅助分类头，以增强训练过程中的梯度流动。 后处理：在SemanticKITTI等数据集上进行密集评估时，需要通过k-NN插值结合多数投票法，将处理后的点的标签传播到全分辨率点云中。我们取k=7，在3-7的典型范围内兼顾效率与抗噪能力[^2] [^13] [^33] [^69]。 损失函数：我们遵循所提出的SAM2-UNet[^80]模型的训练目标，该模型采用多组分损失函数，包括加权交叉熵($L_{WCE}$)、Dice损失($L_{Dice}$)、边界损失($L_{Boundary}$)以及Jaccard指数损失($L_{IoU}$)。这种复合损失函数通过加权交叉熵[^98]解决类别不平衡问题，借助Dice和Jaccard损失[^58] [^71]提升区域级分割精度， ---
Table 1. Comparisons among different Hiera backbones on the test set of SemanticKITTI [^4]. Interestingly, the SAM2-tiny model outperforms its larger counterparts despite having fewer parameters, demonstrating that model capacity does not necessarily correlate with performance in this task domain.
BackboneParams(M)mIoUcarbicymototruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
Hiera-tiny63.361.595.044.347.360.055.268.279.62.694.849.682.11.288.662.085.168.271.364.848.2
Hiera-small70.260.594.747.247.772.239.763.182.50.095.332.180.80.188.360.088.468.678.464.945.5
Table 2. Experiments with and without training augmentations. We observe a strong performance gain when introducing the proposed training augmentations from [^33]. Comparisons among SAM2-tiny on SemanticKITTI test set [^4].
Augs.mIoUcarbicymototruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
off55.093.136.237.253.936.753.759.70.093.435.379.30.085.555.886.459.376.757.145.3
on61.595.044.347.360.055.268.279.62.694.849.682.11.288.662.085.168.271.364.848.2
Table 3. Experiments with and without transfer-learning techniques using Cityscapes dataset [^15]. We observe that overall our model does not benefit from extensive transfer-learning strategies introduced by [^2]. Comparisons among SAM2-tiny on SemanticKITTI [^4] test set.
PretrainmIoUcarbicymototruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
yes59.794.243.238.082.245.356.465.50.094.843.781.90.187.960.786.068.1073.764.346.0
no61.595.044.347.360.055.268.279.62.694.849.682.11.288.662.085.168.271.364.848.2
Table 4. Comparisons among state-of-the-art LiDAR range view semantic segmentation approaches on the validation sequence of SemanticKITTI [^4]. We compare our best performing model using SAM2 with the tiny Hiera backbone. Notably, our model is the first and only approach that leverages VFMs for this task.
MethodVFMmIoUcarbicymototruco.vehpedb.listm.listroadparkwalko.grobuildfencvegtrunterrpolesign
KPRNet[^32] (2020)$\times$63.195.554.147.923.642.665.965.016.593.273.980.630.291.768.485.769.871.258.764.1
LiteHDSeg[^52] (ICRA21)$\times$63.892.340.055.437.739.659.271.654.393.068.278.329.391.565.078.265.865.159.567.7
MPF[^1] (WACV21)$\times$55.593.430.238.326.128.548.146.118.190.662.374.530.688.559.783.559.769.249.758.1
FIDNet[^94] (IROS21)$\times$59.593.954.748.927.623.962.359.823.790.659.175.826.788.960.584.564.469.053.362.8
RangeViT[^82] (ICCV21)$\times$64.095.455.843.529.842.163.958.238.193.170.280.032.592.069.085.370.671.260.864.7
CENet[^14] (ICME22)$\times$64.791.958.650.340.642.368.965.943.590.360.975.131.591.066.284.569.770.061.567.6
MaskRangel[^34] (2022)$\times$66.194.256.055.759.252.467.664.831.891.770.777.129.590.665.284.668.569.260.266.6
RangeFormer[^92] (NeurIPS22)$\times$73.396.769.473.759.966.278.175.958.192.473.078.842.492.370.186.673.372.866.466.6
RangeSAM(Ours)✓60.991.742.044.341.236.155.954.430.292.267.577.228.689.862.884.566.969.858.962.4
并通过边界损失[^74]组件促进精确的边界划分。组合损失函数可表示为加权和形式： $$ \mathcal{L}_{\mathrm{t o t a l}}=\lambda_{1}\mathcal{L}_{W C E}+\lambda_{2}\mathcal{L}_{\mathrm{D i c e}}+\lambda_{3}\mathcal{L}_{\mathrm{B o u n d a r y}}+\lambda_{4}\mathcal{L}_{\mathrm{I o U}}, $$ 其中$\lambda_{i}$为根据经验确定的权重系数。在我们的实验配置中，设定$\lambda_{i}=1$，赋予各损失组分同等重要性。 #### 4.1. 数据集选择与评估指标 我们以SemanticKITTI[^4]和nuScenes[^8]作为主要训练数据集。SemanticKITTI包含用64线激光雷达采集的城市驾驶场景数据，分为19,130个训练扫描序列（序列00-07、09-10），4,071个… 接下来我们将介绍所使用的数据集、采用的训练配置以及评估指标。 ### 4. 数据集与实现细节 验证扫描（序列08），以及涵盖19个语义类别的20,351次测试扫描（序列11-21）。NuScenes数据集包含来自波士顿和新加坡的1,000个城市场景，这些场景由32线激光雷达采集，提供了16个类别下的28,130帧训练数据和6,019帧验证数据。为了保持一致，我们将NuScenes转换为SemanticKITTI格式，并在SemanticKITTI验证集上报告平均交并比（mIoU）[^19]。 #### 4.2. 数据增强 我们采用了标准的增强方法，包括全局旋转、坐标抖动、翻转以及随机点消除（所有操作的概率均为1.0），同时还应用了源自[^33]的范围视图专用增强技术，即混合、合并、平移和复制粘贴，它们的概率分别为[0.9, 0.1, 0.9, 1.0]。 #### 4.3. 训练策略 所有实验均使用8块A100 GPU（40GB），每块GPU的批量大小为2（有效批量大小为16）。训练过程中… --- 采用了为期5个周期的线性热身阶段，随后进行余弦退火调节[^23]。鉴于近期研究[^2] [^33] [^77] [^78]表明基于Transformer的架构能从预训练中受益，我们采用了[^2] [^33]提出的训练方案，在NuScenes数据集[^8]（已转换为SemanticKITTI格式）上预训练60个周期。我们选择SAM2-tiny作为骨干网络，因为更大型的SAM2变体带来的性能提升微乎其微（见表1），却会大幅增加计算成本和推理延迟。最终模型包含约6300万个参数。优化方面：我们采用AdamW优化器，并设置差异化的学习率：Hiera骨干网络的学习率为0.0004，权重衰减为$\alpha = 0.001$；模型其余部分的学习率为0.001，权重衰减为$\alpha = 0.0001$。 #### 4.4. 评估 我们展示了RangeSAM在SemanticKITTI数据集上的实验结果[^4]。表4显示了序列08上的验证结果，而表1、表2和表3则报告了测试集上的性能表现。与现有最先进方法相比，我们的方法取得了具有竞争力的成果。 ![Image](images/7de66502cff42a9decfe6cebb0ea2894602a7934cf1b3b11b8794bdc76464f8e.jpg)
(a)
![Image](images/b04c9ca6b79149e7ca7346a85aaeda8c85c70de48796ed637aa0b303b2e5ceb0.jpg)
(b)
![Image](images/1a3fae19dfa65fa99c1d051219031f6bc6dc8da1c21120fd27c63ca3cd22c250.jpg)
(c)
Figure 2. Qualitative segmentation examples of increasing difficulty using our model. (a) Urban intersection, (b) Suburban environment and (c) Highly cluttered scene.
#### 4.5. 定量评估 表1对比了SAM2 Hiera骨干网络的不同变体：tiny版和small版在mIoU指标上表现相近，其中tiny版凭借更少的参数和更快的推理速度具备计算优势。 表4将我们的模型与基线模型进行了对比。尽管近期最先进方法的mIoU能达到65%–75%，但我们的模型仍达到了60.9%，成功整合了SAM2视觉基础模型[^50]。性能可分为三个层级： • 高IoU（80%–90%）：大型且出现频率高的类别（汽车、道路、建筑物、植被）的性能与最先进方法相当。 • 中等IoU（60%–70%）：中等频率的类别（卡车、围栏、地形）仍具备竞争力。 - 低IoU（29%–47%）：稀有且体积较小的物体（摩托车、自行车、行人）识别难度较大，这一结果与现有方法的表现一致。 虽然长尾类别的性能较低，反映出相对于6300万个参数而言训练样本数量有限，但在表4中标注的代表性类别上，我们的结果仍具竞争力。通过扩展数据集和优化设计，有望缩小与长尾类别之间的差距。 #### 4.6. 定性评估 图2展示了三种分割场景：(a) 城市交叉口：模型能够准确分割车辆、静态元素（树木、墙壁、围栏）以及细小结构（围栏柱）。(b) 郊区环境：主要类别（汽车、植被、围栏）被正确分割，但模型对输电塔等未见过的类别识别能力较弱。(c) 高度杂乱的场景：尽管物体相互重叠（树木、围栏、电线杆、标志牌），主要类别仍能被有效识别。较小或距离较远的结构会存在噪声，但整体分割效果依然稳健。 #### 4.7. 消融实验 我们开展了全面的消融实验，旨在系统评估不同训练策略对模型性能的影响。 迁移学习：我们采用了该研究领域其他基线方法所采用的迁移学习思路$^{[2]}$。在表3对应的实验中，我们先在Cityscapes数据集[^15]上对模型进行25个epoch的预训练，以便在目标数据集上训练前强化特征学习能力。然而，这种方法反而导致了性能下降。我们认为这可能是由于SAM2模型已在大规模图像数据集上经过了充分预训练，从而产生了域差异，影响了迁移学习效果。要证实这一假设并探索其他适配策略，还需进一步研究。 距离视图增强：结果（见表2）表明，将[^33]引入的训练流程中的增强手段，几乎能提升所有数据集上的性能，mIoU指标可提升10%。 ### 5. 结论 据我们所知，本研究是首个将视觉基础模型应用于距离视图激光雷达点云分割的工作。尽管RGB图像与距离视图表示之间存在域差异，但我们对SAM2编码器和解码器进行的针对性架构修改有效弥合了这种模态差异，最终取得了具有竞争力的性能。我们整合了以往研究中的增强策略和训练方案[^2] [^13] [^33] [^77]。值得注意的是，我们发现跨多种3D基准测试集的多数据集训练效果，优于如[^2]中所述的大规模2D预训练。我们将发布源代码和模型权重。 未来工作：主要局限在于计算复杂度。无卷积结构的Hiera骨干网络[^55]在处理稀疏点云时表现不佳，需要借助RFB[^37]才能取得有竞争力的性能。不过，RFB构成了主要的计算瓶颈，目前阻碍了其实时部署——这也是未来优化的关键目标。 --- ## 参考文献 [^1]: Yara Ali Alnaggar, Mohamed Afifi, Karim Amer, and Mohamed Elhelw. Multi-Projection Fusion for Real-Time Semantic Segmentation of 3D LiDAR Point Clouds. In Winter Conference on Applications of Computer Vision (WACV), 2021. 4 [^2]: Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5240–5250, 2023. 1, 2, 3, 4, 5 [^3]: Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization, 2016. 2, 3 [^4]: J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV), 2019. 2, 4, 5 [^5]: J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, J. Gall, and C. Stachniss. Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset. The International Journal on Robotics Research, 40(8-9):959–967, 2021. 2 [^6]: Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren E. Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 2 [^7]: Nhat-Tan Bui, Dinh-Hieu Hoang, Minh-Triet Tran, Gianfranco Doretto, Donald Adjeroh, Brijesh Patel, Arabinda Choudhary, and Ngan Le. Sam3d: Segment anything model in volumetric medical images, 2024. 2 [^8]: Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. CoRR, abs/1903.11027, 2019. 2, 4, 5 [^9]: Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 2 [^10]: Jiale Chen, Fei Xia, Jianliang Mao, Haoping Wang, and Chuanlin Zhang. Siesef-fusionnet: Spatial inter-correlation 用于激光雷达点云语义分割的增强与空间嵌入特征融合网络，2024年。2 [^11]: Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip, 2023. 2 [^12]: Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more, 2024. 2 [^13]: Tzu-Hsuan Chen and Tian Sheuan Chang. Rangeseg: Range-aware real time segmentation of 3d lidar point clouds. IEEE Transactions on Intelligent Vehicles, 7(1):93–101, 2022. 1, 2, 3, 5 [^14]: Hui-Xian Cheng, Xian-Feng Han, and Guo-Qiang Xiao. Cenet: Toward concise and efficient lidar semantic segmentation for autonomous driving. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 01–06. IEEE, 2022. 4 [^15]: Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 4, 5 [^16]: Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds for autonomous driving, 2020. 2 [^17]: Aashish Dhawan, Pankaj Bodani, and Vishal Garg. Post processing of image segmentation using conditional random fields. In 2019 6th International Conference on Computing for Sustainable Global Development (INDIACom), pages 729–734, 2019. 2 [^18]: Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi proposal aggregation for 3d semantic instance segmentation. CoRR, abs/2003.13867, 2020. 1 [^19]: Mark Everingham, S. M. Eslami, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. Int. J. Comput. Vision, 111(1):98–136, 2015. 4 [^20]: Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022. 2 [^21]: Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 2 [^22]: Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354–3361, 2012. 3 ## [^23]: Akhilesh Gotmare, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. A closer look at deep learning heurism --- 技巧：学习率重启、热身与蒸馏。CoRR, abs/1810.13243, 2018年。5 [^24]: Yi Gu, Yuming Huang, Chengzhong Xu, and Hui Kong. Maskrange: A mask-classification model for range-view based lidar segmentation, 2022. Under review. 4 [^25]: Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 2, 3 [^26]: Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation, 2022. 1 [^27]: Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11105–11114, Los Alamitos, CA, USA, 2020. IEEE Computer Society. 2 [^28]: Guohao Huo, Ruiting Dai, and Hao Tang. Samba-unet: Synergizing sam2 and mamba in unet with heterogeneous aggregation for cardiac mri segmentation, 2025. 2 [^29]: Sergey Ioffe and Christian Szegedy. Batch normalization: accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, page 448–456. JMLR.org, 2015. 3 [^30]: Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything, 2023. 2 [^31]: Deyvid Kochanov, F. Karimi Nejadasl, and Olaf Booij. Kprnet: Improving projection-based lidar semantic segmentation. ArXiv, abs/2007.12668, 2020. 2 [^32]: Deyvid Kochanov, Fatemeh Karimi Nejadasl, and Olaf Booij. KPRNet: Improving projection-based LiDAR semantic segmentation. arXiv preprint arXiv:2007.12668, 2020.4 [^33]: Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation, 2023. 1, 2, 3, 4, 5 [^34]: Loïc Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. CoRR, abs/1711.09869, 2017. 2 [^35]: Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. CoRR, abs/1605.07648, 2016. 3 [^36]: Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 2 [^37]: Songtao Liu, Di Huang, and Yunhong Wang. Receptive field block net for accurate and fast object detection. CoRR, abs/1711.07767, 2017. 1, 2, 3, 5 [^38]: Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded 用于开放集物体检测的预训练。arXiv预印本arXiv:2303.05499，2023年。2 [^39]: Yuyuan Liu, Yuanhong Chen, Chong Wang, Junlin Han, Junde Wu, Can Peng, Jingkun Chen, Yu Tian, and Gustavo Carneiro. Auralsam2: Enabling sam2 hear through pyramid audio-visual feature prompting, 2025. 2 [^40]: Liane-Marina Messmer, Christoph Reich, and Djaffar Ould Abdeslam. Context-aware machine learning: A survey. In Proceedings of the Future Technologies Conference (FTC) 2024, Volume 1, pages 252–272, Cham, 2024. Springer Nature Switzerland. 1 [^41]: Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyrill Stachniss. Rangenet ++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4213–4220, 2019. 1, 2 [^42]: Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Q. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russ Howes, Po-Yao (Bernie) Huang, Shang-Wen Li, Ishan Misra, Michael G. Rabbat, Vasu Sharma, Gabriel Synnaeve, Huijiao Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. ArXiv, abs/2304.07193, 2023.2 [^43]: Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7459–7468, 2021. 1 [^44]: Yancheng Pan, Biao Gao, Jilin Mei, Sibo Geng, Chengkun Li, and Huijing Zhao. Semanticposs: A point cloud dataset with large quantity of dynamic instances. CoRR, abs/2002.09147, 2020. 2 [^45]: Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 815–824, 2023. 2 [^46]: Quang-Hieu Pham, Duc Thanh Nguyen, Binh-Son Hua, Gemma Roig, and Sai-Kit Yeung. JSIS3D: joint semantic-instance segmentation of 3d point clouds with multi-task pointwise networks and multi-value conditional random fields. CoRR, abs/1904.00699, 2019. 1 [^47]: Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. CoRR, abs/1612.00593, 2016. 1 [^48]: Charles Ruizhongtai Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detection from RGB-D data. CoRR, abs/1711.08488, 2017. 1 [^49]: Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. CoRR, abs/1706.02413, 2017. 1 ## [^50]: Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt --- 潘宁、Kalyan Vasudev Alwala、Nicolas Carion、吴超远、Ross Girshick、Piotr Dollár和Christoph Feichtenhofer。Sam 2：图像与视频中的任意物体分割，2024年。1, 2, 3, 5 [^51]: Ryan Razani, Ran Cheng, Ehsan Taghavi, and Liu Bingbing. Lite-hdseg: Lidar semantic segmentation using lite harmonic dense convolutions. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 9550–9556, 2021. 2 [^52]: Ryan Razani, Ran Cheng, Ehsan Taghavi, and Liu Bing-bing. Lite-HDSeg: LiDAR Semantic Segmentation Using Lite Harmonic Dense Convolutions. arXiv preprint arXiv:2103.08852, 2021. 4 [^53]: Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2 [^54]: Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 2 [^55]: Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, Jitendra Malik, Yanghao Li, and Christoph Feichtenhofer. Hiera: a hierarchical vision transformer without the bells-and-whistles. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 2, 5 [^56]: Mohsen Shahraki, Ahmed Elamin, and Ahmed El-Rabbany. Samnet++: A segment anything model for supervised 3d point cloud semantic segmentation. Remote Sensing, 17(7), 2025. 2 [^57]: Oriane Siméoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisao, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Hervé Jégou, Patrick Labatut, and Piotr Bojanowski. Dinov3, 2025. 2 [^58]: Carole H. Sudre, Wenqi Li, Tom Vercauteren, Sébastien Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. CoRR, abs/1707.03237, 2017. 3 [^59]: Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 [^60]: Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2 [^61]: Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. Eva-clip18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2023. 2 [^62]: Zhiyu Tan, Zichao Dong, Cheng Zhang, Weikun Zhang, Hang Ji, and Hao Li. Ovo: Open-vocabulary occupancy, 2023. 2 [^63]: Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. CoRR, abs/2007.16100, 2020. 1 [^64]: Hugues Thomas, C. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6410–6419, 2019. 1, 2 [^65]: Larissa T. Triess, David Peter, Christoph B. Rist, and J. Marius Zöllner. Scan-based semantic segmentation of lidar point clouds: An experimental study. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1116–1121, 2020. 2 [^66]: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. CoRR, abs/1706.03762, 2017. 1 [^67]: Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks, 2018. 1 [^68]: Bernie Wang, Virginia Wu, Bichen Wu, and Kurt Keutzer. Latte: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 265–272. IEEE, 2019. 2 [^69]: Song Wang, Jianke Zhu, and Ruixiang Zhang. Meta-rangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. IEEE Robotics and Automation Letters, 7(4):9739–9746, 2022. 2, 3 [^70]: Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. CoRR, abs/1801.07829, 2018. 1 [^71]: Zifu Wang, Xuefei Ning, and Matthew B. Blaschko. Jaccard metric losses: Optimizing the jaccard index with soft labels, 2024. 3 [^72]: Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. 2 [^73]: Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, 2019. 2 [^74]: Dongyue Wu, Zilin Guo, Aoyan Li, Changqian Yu, Changxin Gao, and Nong Sang. Conditional boundary loss for semantic segmentation. IEEE Transactions on Image Processing, 32:3717–3731, 2023. 4 ## [^75]: Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In 2019 --- IEEE/CVF计算机视觉与模式识别会议（CVPR），第9613–9622页，2019年。1 [^76]: Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Heng-shuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1 [^77]: Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1, 5 [^78]: Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 5 [^79]: Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 2 [^80]: Xinyu Xiong, Zihuang Wu, Shuangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv preprint arXiv:2408.08870, 2024. 2, 3 [^81]: Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1–19. Springer, 2020. 2 [^82]: Chenfeng Xu, Bichen Wu, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Rangevit: Vision transformer for lidar range image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 27069–27078, 2021. 4 [^83]: Jianyun Xu, Song Wang, Ziqian Ni, Chunyong Hu, Sheng Yang, Jianke Zhu, and Qiang Li. Sam4d: Segment anything in camera and lidar streams, 2025. 2 [^84]: Xiang Xu, Lingdong Kong, Hui Shuai, and Qingshan Liu. Frnet: Frustum-range networks for scalable lidar segmentation. IEEE Transactions on Image Processing, 34:2173–2186, 2025. 2 [^85]: Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5588–5597, 2020. 1 [^86]: Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes, 2023. 2 [^87]: Haobo Yuan, Xiangtai Li, Tao Zhang, Zilong Huang, Shilin Xu, Shunping Ji, Yunhai Tong, Lu Qi, Jiashi Feng, and Ming-Hsuan Yang. Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos, 2025. 2 [^88]: Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and 张鹏川。Florence：一种全新的计算机视觉基础模型。CoRR，abs/2111.11432，2021年。2 [^89]: Karim Abou Zeid, Kadir Yilmaz, Daan de Geus, Alexander Hermans, David Adrian, Timm Linder, and Bastian Leibe. Dino in the room: Leveraging 2d foundation models for 3d segmentation, 2025. 2 [^90]: Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip. In 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 2040–2051, 2023. 2 [^91]: Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9598–9607, 2020. 2 [^92]: Zhen Zhang, Yin Zhou, Hassan Foroosh, Mikko Lauri, Shiguang Li, Hongkai Xu, and Vincent Lepetit. Rangeformer: Point cloud semantic segmentation using range image based transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 22397–22410, 2022. 4 [^93]: Rongzhen Zhao, Vivienne Wang, Juho Kannala, and Joni Pajarinen. Multi-scale fusion for object representation, 2025. 1 [^94]: Yiming Zhao, Lin Bai, and Xinming Huang. FIDNet: LiDAR Point Cloud Semantic Segmentation with Fully Interpolation Decoding. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 4 [^95]: Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. CoRR, abs/2008.01550, 2020. 1 [^96]: Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16260–16270, 2021. 2 [^97]: Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once, 2023. 2 [^98]: Ozgür Ozdemir and Elena Battini Sönmez. Weighted cross-entropy for unbalanced data with application on covid x-ray images. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 1–6, 2020. 3