Research map / Point clouds / Point cloud classification
TOD3Cap: Towards 3D Dense Captioning in Outdoor Scenes
Bu Jin, Yupeng Zheng, Pengfei Li, Weize Li, Yuhang Zheng, Sujie Hu, Xinyu Liu, Jinwei Zhu, Zhijie Yan, Haiyang Sun, Kun Zhan, Peng Jia and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Relation Q-Former instead of relational graph (ablation) | TOD3Cap network · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 82.7 -> 94.1 | Improved |
| use Relation Q-Former instead of transformer decoder (ablation) | TOD3Cap network · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 90.0 -> 94.1 | Improved |
| use LLaMA-Adapter decoder instead of GPT2 (ablation) | TOD3Cap network · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 85.6 -> 94.1 | Improved |
| use LLaMA instead of S&T decoder (ablation) | TOD3Cap network · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 78.6 -> 94.1 | Improved |
| add detector pretraining stage (ablation) | TOD3Cap network training · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 69.5 -> 94.1 | Improved |
| add captioner pretraining stage (ablation) | TOD3Cap network training · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 85.3 -> 94.1 | Improved |
| add camera images to LiDAR-only input (ablation) | TOD3Cap network · TOD3Cap val | CIDEr@0.5IoU: 74.4 -> 108.0 | Improved |
| add LiDAR to camera-only input (ablation) | TOD3Cap network · TOD3Cap val | CIDEr@0.5IoU: 94.1 -> 108.0 | Improved |
| use larger BEV resolution and backbone (Small to full) (ablation) | TOD3Cap network · TOD3Cap val (2D camera input) | CIDEr@0.5IoU: 87.5 -> 94.1 | Improved |
| adapt indoor baselines with the BEV detector | Scan2Cap · TOD3Cap | training: training collapse -> C@0.5 62.5 (2D+3D) | Improved |
| add LLM-generated driving instructions as input to UniAD | UniAD · nuScenes planning | Avg. collision: 0.31 -> 0.23 | Improved |
| add LLM-generated driving instructions as input to UniAD | UniAD · nuScenes planning | Avg. L2: 1.03 -> 1.05 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Scan2Cap* — Scan2Cap: Context-Aware Dense Captioning in RGB-D Scans (CVPR 2021) · better on 2 of 2 shared results
- X-Trans2Cap* — X-Trans2Cap: Cross-Modal Knowledge Transfer Using Transformer for 3D Dense Captioning (CVPR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Vote2Cap-DETR++: Decoupling Localization and Describing for End-to-End 3D Dense Captioning (TPAMI 2024 · similarity 0.91)
- Scan2Cap: Context-Aware Dense Captioning in RGB-D Scans (CVPR 2021 · similarity 0.89)
- LiDAR-LLM: Exploring the Potential of Large Language Models for 3D LiDAR Understanding (AAAI 2025 · similarity 0.89)
- Dense Video Object Captioning from Disjoint Supervision (ICLR 2025 · similarity 0.87)
- MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations (NeurIPS 2024 · similarity 0.87)
- USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
