Research map / 3D vision / 3D generation

Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding

CVPR 2025 · 3D vision › 3D generation · approach: Learning 3D reconstruction from images · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TSP3D, a fully sparse multi-level convolutional single-stage 3D visual grounding model that uses text-guided voxel pruning and completion-based addition to run real-time while matching or beating two-stage accuracy.

Problem

Single-stage and two-stage 3D visual grounding methods are too slow for real-time use, and point-based backbones are costly; sparse convolution backbones have too many voxels for cross-modal attention with text.

Approach

Text-guided pruning (TGP) sparsifies the multi-level voxel features using text-conditioned cross-attention so that expensive voxel-text interaction becomes affordable, and completion-based addition (CBA) recovers over-pruned target voxels from high-resolution backbone features.

Key results

• On ScanRefer, TSP3D reaches 56.45 Acc@0.25 and 46.71 Acc@0.5 at 12.43 FPS, +1.13 Acc@0.5 over the best two-stage method (MCLN, 45.53) and about 2x the FPS of prior single-stage methods.
• On Nr3D and Sr3D in the single-stage setting, TSP3D reaches 48.7% and 57.1%, leading the best two-stage (detected boxes) results by +2.6 and +3.2 points.

Measured effects reported in the paper

Surprising findings

• Full cross-attention over the dense voxel set exceeds GPU memory, so attention-based fusion was infeasible without pruning.
• CBA alone without TGP gives little gain because no voxels are pruned.

Limitations

• Single-stage speed gains are reported relative to reproduced baselines, and some methods' speeds were measured by the authors.
• Pruning may remove small or thin targets, which CBA only partly recovers.

Experimental setup

Datasets: ScanRefer Nr3D Sr3D

Benchmarks: ScanRefer Nr3D Sr3D

Metrics: Acc@0.25 Acc@0.5 inference speed (FPS)

Baselines: ScanRefer TGNN InstanceRefer SAT FFL-3DOG 3D-SPS BUTD-DETR EDA 3D-VisTA VPP-Net G3-LQ MCLN

Builds on: FCAF3D TR3D GeSpConv (generative sparse convolution) BUTD-DETR MinkowskiEngine

Closest papers by meaning

• SimVG: A Simple Framework for Visual Grounding with Decoupled Multi-modal Fusion (NeurIPS 2024 · similarity 0.87)
• Mono3DVG: 3D Visual Grounding in Monocular Images (AAAI 2024 · similarity 0.87)
• G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding (CVPR 2024 · similarity 0.87)
• CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large Vision-Language Models (ICML 2026 · similarity 0.87)
• SPAZER: Spatial-Semantic Progressive Reasoning Agent for Zero-shot 3D Visual Grounding (NeurIPS 2025 · similarity 0.86)
• Towards CLIP-driven Language-free 3D Visual Grounding via 2D-3D Relational Enhancement and Consistency (CVPR 2024 · similarity 0.86)