Parsing · 读懂
拆句练习
论文的长句不是靠「读得快」,是靠先找到主干。60 句按难度排好,从 L1 开始。
已读懂 0 / 60 句
0 句试过
L1
L1 · 一个从句
19 句20–25 词。先找主谓宾,剩下的都是修饰。
002 Rather than directly predicting absolute target features, the network learns a precise residual correction to filter out uncalibrated collaborative noise.FRUC · 3.3 Cross-Agent Latent Residual Denoising → 003 In the zero-shot scenario, FRUC consistently outperforms feed-forward baselines, demonstrating robust cross-dataset adaptability and reliable structural priors learned from V2X-Real.FRUC · 4.2 Benchmark Comparison → 005 The intuition is that while collaborative features are essential for completing blind spots, they may introduce uncalibrated geometric drift into visible regions.FRUC · 3.3 Cross-Agent Latent Residual Denoising → 008 Existing 3D distillation methods are not designed for this setting, where long-range depth accuracy and multi-view spatial relations must be preserved under large model compression.LiAuto-GeoX · 2 Related Work → 009 However, such point-wise alignment ignores the relational structure among views, which is crucial for constructing a coherent 3D driving scene.LiAuto-GeoX · 3.2.2 Relative-Pose Relational Distillation → 012 However, most existing geometry reconstruction methods (e.g., DVGT) rely on computationally expensive batch processing of multi-frame inputs and cannot be applied to online planning.DVGT-2 · Abstract → 016 Second, similar to the global pointmap evaluation, deriving global ego-poses by accumulating relative poses inevitably introduces trajectory drift over time.DVGT-2 · 4.4 Geometry Reconstruction → 020 Compared to the base VLM, VGGDrive achieves a 2.72 improvement in PDMS, while VGGT-Dist and VGGT-Add show only marginal performance gains.VGGDrive · 4.3. Performance Comparison → 023 Building photorealistic reconstructions of dynamic urban scenes is crucial for autonomous driving, as it supports closed-loop training and evaluation under realistic motion patterns.DynamicVGGT · 2.2. 3DGS Reconstruction for Driving Scenes → 024 This motivates the need for a unified feed-forward framework that can jointly model geometry and motion, enabling temporally consistent dynamic scene reconstruction.DynamicVGGT · 1. Introduction → 027 We observe that freezing the AA blocks causes the pretrained VGGT backbone to overemphasize geometric reasoning while weakening appearance cues, which degrades Gaussian rendering quality.DynamicVGGT · 3.4. Dynamic 3D Gaussian Splatting Head → 030 Next, a spatial projector is utilized to map the latent BEV features into the language space and obtain vision tokens.DrivePI · 3.1. Overview → 031 These results verify the effectiveness of unifying text understanding with 3D perception, prediction, and planning tasks within a single VLA framework.DrivePI · 4.4. Ablation Study → 033 However, most existing approaches realize this reasoning at the text level, leading to high latency and exposure bias from autoregressive decoding.ColaVLA · 3.2. Cognitive Latent Reasoning → 035 Among action-based approaches, ColaVLA attains the best overall accuracy and safety, achieving the lowest average L2 error (0.30m) and the lowest average collision rate (0.23%).ColaVLA · 4.2. Comparison with State-of-the-art Methods → 036 Experiments on the nuScenes benchmark show that ColaVLA achieves state-of-the-art performance in both open-loop and closed-loop settings with favorable efficiency and robustness.ColaVLA · Abstract → 043 A key future direction is to extend our causal framework from the sequential paradigm of VAD to parallel or iterative interaction architectures.CausalVAD · 5. Conclusion → 044 The above analysis reveals that causal confounding in the VAD framework is not a single failure point but a cascading systemic issue.CausalVAD · 3.1. Causal graph and problem formulation → 045 Instead, we advocate for a fundamental reshaping of the model’s learning objective: shifting from fitting correlations to inferring causation.CausalVAD · 1. Introduction →
L2
L2 · 两个从句或分词
28 句25–35 词。注意逗号后面的 -ing / -ed 部分。
001 To achieve non-destructive geometric supplementation under uncalibrated cross-agent misalignment, FRUC first introduces an ego-centric causal occlusion field that explicitly derives occlusion evolution as latent priors by modeling agent-wise spatio-temporal correlations.FRUC · Abstract → 004 Furthermore, despite incorporating advanced spatial reasoning modules to handle uncalibrated collaborative context, our framework maintains a highly competitive inference speed of 0.77s, significantly faster than optimization-based paradigms.FRUC · 4.2 Benchmark Comparison → 006 Existing large-scale visual geometry models typically require substantial computational resources and lack the long-range geometric fidelity, surround-view consistency, and real-time efficiency demanded by dynamic driving environments.LiAuto-GeoX · Abstract → 007 This framework employs mask-guided depth-aware distillation to retain fine-grained metric structures by emphasizing geometrically informative regions, and relative-pose relational distillation to enforce cross-view spatial consistency through pose-induced geometric relations.LiAuto-GeoX · Abstract → 010 During inference, the teacher, LiDAR inputs, and mask adapter are removed; only the compact student is deployed to produce dense 3D reconstruction in a single forward pass.LiAuto-GeoX · 3.3 Training Procedure and Loss Formulation → 011 While large models still achieve stronger accuracy on some datasets, such as Waymo and PandaSet, LiAuto-GeoX maintains comparable point-cloud reconstruction quality under a much smaller model scale.LiAuto-GeoX · 4.1 Surround-View 3D Reconstruction → 013 To overcome the computational bottleneck of traditional multi-frame processing, we propose a sliding-window streaming strategy with temporal causal attention and feature caching, enabling efficient, on-the-fly joint prediction of geometry and trajectories.DVGT-2 · 5 Conclusion → 014 However, DVGT-2 achieves a stable latency of around 260ms per frame, even across hundreds of frames, confirming that our model is well-suited for real-time, infinite-length autonomous driving.DVGT-2 · 4.6 Inference Efficiency Comparison → 015 While this enables efficient per-frame processing, it lacks the global context utilized by baseline methods that access the entire sequence simultaneously, thereby limiting long-term global ego-pose consistency.DVGT-2 · 4.4 Geometry Reconstruction → 017 Although language descriptions offer valuable high-level context, they can sometimes be too coarse to capture complete and accurate geometric details when precise spatial control is required.DVGT-2 · 1 Introduction → 018 The significance of cross-view 3D geometric modeling capabilities for autonomous driving is self-evident, yet existing Vision-Language Models (VLMs) inherently lack this capability, resulting in their mediocre performance.VGGDrive · Abstract → 019 Concretely, to bridge the cross-view 3D geometric features from the frozen visual 3D model with the VLM’s 2D visual features, we introduce a plug-and-play Cross-View 3D Geometric Enabler (CVGE).VGGDrive · Abstract → 021 While this approach improves trajectory performance, it disconnects scene understanding from decision-making, preventing the model’s knowledge and reasoning from being effectively translated into final control outputs.VGGDrive · 1. Introduction → 022 We thus chart a new course, advocating for the infusion of VLMs with the cross-view geometric grounding of mature 3D foundation models, closing this critical capability gap in autonomous driving.VGGDrive · 1. Introduction → 025 Under the multi-view stereo setting, DynamicVGGT achieves the best overall results with an Abs Rel of 0.051 and 97.6% accuracy, surpassing VGGT and StreamVGGT by a clear margin.DynamicVGGT · 4.5. Monocular and MVS depth estimation → 026 On the KITTI dataset, which uses monocular input with three consecutive frames per sequence, DynamicVGGT achieves the best results across most metrics, obtaining an accuracy of 0.901 and a normal consistency of 0.939.DynamicVGGT · 4.3. Point Map Reconstruction → 028 Remarkably, with only a 0.5B Qwen2.5 model as MLLM backbone, DrivePI as a single unified model matches or exceeds both existing VLA models and specialized VA models.DrivePI · Abstract → 029 However, existing LLM-based methods struggle to directly output fine-grained perception results (e.g., 3D occupancy and occupancy flow) through next-token prediction, which are readily achievable by VA approaches.DrivePI · 3. Methods → 032 First, we use a simple multi-task learning strategy to balance loss weights across tasks, which may not be optimal to obtain the best trade-off between competing objectives.DrivePI · 5. Conclusion → 034 During training, the selection is made differentiable using a Gumbel–Softmax relaxation to form a K-hot mask, while at inference time the top-K tokens are retained directly.ColaVLA · 3.2. Cognitive Latent Reasoning → 037 Despite progress, most E2E methods rely on sparse supervision from a single ground truth trajectory, which limits reasoning and diversity, highlighting the need for higher-level semantic and cognitive guidance.ColaVLA · 2.1. End-to-end Autonomous Driving → 038 A significant bottleneck in this approach is its substantial demand for high-quality labeled data, such as 3D bounding boxes and semantic segmentation, which are especially expensive to annotate manually.ActiveAD · Abstract → 039 An astonishing finding is that ActiveAD achieves comparable or even better performance by utilizing a carefully selected 30% of the data compared to training with the entire 100% dataset.ActiveAD · 5.2. Performance by Planning Metrics → 040 Firstly, data in AD often involves rich multi-modality information, such as video streams, driving trajectories, and various meta-information like vehicle speed, whereas most existing active learning methods typically consider only single-modal images as input.ActiveAD · 1. Introduction → 041 SCIS operationalizes the do-operator within the neural network by leveraging a pre-computed confounder dictionary to represent latent driving contexts and performing a parameterized causal intervention on the efficient vectorized scene queries of the VAD model.CausalVAD · 1. Introduction → 042 This significant performance gain, especially the dramatic improvement in safety, strongly indicates that by introducing causal intervention, the model learns to better understand scene dynamics and make more reliable decisions.CausalVAD · 4.2.1. Open-loop planning performance → 054 Unlike existing approaches that treat VA and VLA-based methods as separate paradigms, DrivePI establishes a unified architecture that seamlessly integrates the spatial precision of VA models with the interpretability and interactive capabilities of VLA frameworks.DrivePI · 1. Introduction → 056 At a high level, ColaVLA consists of a Cognitive Latent Reasoner that distills scene evidence into compact meta-action priors, and a Hierarchical Parallel Planner that converts these priors to multi-scale trajectories under a causality-preserving scheme.ColaVLA · 1. Introduction →
L3
L3 · 从句套从句
13 句35–50 词。先拆主干,再把从句一层层挂回去。
046 In this paper, we rethink this task by conceptualizing a distributed multi-vehicle network as a spatio-temporally unstructured ego-centric multi-camera system, where the core challenge lies in enhancing ego-centric occluded geometry through collaboration without degrading the ego’s accurately observed visible geometry, while preserving reconstruction efficiency.FRUC · Abstract → 047 Forcing a network to jointly encode this highly irregular context creates a cluttered latent space, which overwhelms the network and severely corrupts the ego vehicle’s previously established geometric priors, even destroying its reliable single-vehicle reconstruction capabilities (e.g., ghosting or distorted structures).FRUC · 1 Introduction → 048 Through extensive evaluations on real-world V2X-Real and UrbanIng-V2X datasets, FRUC is shown to be a new state-of-the-art for the scene reconstruction of dynamic collaborative driving environments, significantly outperforming existing methods in both rendering quality and efficiency.FRUC · Abstract → 049 Because the backbone is kept strictly frozen without additional fine-tuning or reinforcement learning, this setting properly evaluates whether reconstructed 3D geometry can serve as a transferable representation for planning, rather than relying on task-specific adaptation of the entire network.LiAuto-GeoX · 4.5 Downstream Task Comparison → 050 Compared with larger generic visual geometry models, which can still exhibit fragmented layouts, inconsistent cross-view geometry, or missing distant structures, our method better preserves the spatial arrangement of roads, vehicles, and scene boundaries in many driving cases.LiAuto-GeoX · 4.1 Surround-View 3D Reconstruction → 051 Unlike traditional end-to-end models that rely on multi-modal inputs and sparse perceptual supervision, or VLA models that require language labels and complicated RL fine-tuning, our VGA paradigm achieves safe, end-to-end driving using only annotation-efficient geometric supervision.DVGT-2 · 4.5 Planning → 052 To effectively bridge the 3D geometric features with the 2D visual representations in the LLM and enable the 3D features to deeply steer the VLM’s performance in specific driving tasks, a plug-and-play Cross-view 3D Geometric Enabler (CVGE) is proposed.VGGDrive · 1. Introduction → 053 The CVGE decouples the base LLM architecture and introduces a hierarchical adaptive injection mechanism, which injects 3D features into the LLM’s 2D visual embeddings in a multilevel, structured manner, thereby establishing genuine geometric grounding within the model.VGGDrive · 1. Introduction → 055 Second, DrivePI generates intermediate fine-grained 3D perception (e.g., 3D occupancy) and prediction (e.g., occupancy flow) representations to ensure that the output features of the MLLM maintain reliable spatial perception capabilities, thereby enhancing both interpretability and safety guarantees for autonomous driving systems.DrivePI · 1. Introduction → 057 This abstraction subsumes three dominant paradigms in the field: (1) classical modular pipelines with separate perception, decision, and planning stages; (2) end-to-end planners without explicit reasoning; and (3) VLM-based models that formulate planning as autoregressive text generation with chain-of-thought reasoning.ColaVLA · 3.1. Framework Overview → 058 While there is still a gap compared to models specifically optimized for closed-loop performance like DriveMoE [35], our results clearly corroborate that by learning more fundamental causal relationships, the model acquires stronger generalization, enabling it to effectively transfer its driving knowledge to unseen scenarios.CausalVAD · 4.2.2. Analysis of causal robustness → 059 Even worse, it is widely acknowledged that the AD task has serious long-tailed issues [23], which means a large part of collected data is trivial e.g. simply driving forward in a straight road, and only a few cases are safety-critical.ActiveAD · 1. Introduction → 060 The patch and camera tokens are processed by the Alternating-Attention (AA) blocks to model intra-frame spatial geometry, whereas the Motion-aware Temporal Attention (MTA) blocks operate in parallel to model inter-frame temporal dependencies using the motion tokens.DynamicVGGT · 2. Related work →