Research map / Video / Video-language understanding
4D-Bench: Benchmarking Multi-modal Large Language Models for 4D Object Understanding
Wenxuan Zhu, Bing Li, Cheng Zheng, Jinjie Mai, Jun Chen, Letian Jiang, Abdullah Hamdi, Sara Rojas Martinez, Chia-Wen Lin, Mohamed Elhoseiny, Bernard Ghanem
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 |
|---|---|---|---|
| increase number of input views (ablation) | Gemini 1.5 Flash · 4D-Bench 4D object QA | accuracy (%): 41.3 -> 53.7 | Improved |
| increase temporal sampling frequency (ablation) | Gemini 1.5 Flash · 4D-Bench 4D object QA | accuracy (%): 46.3 -> 53.7 | Improved |
| use more than 3 views or 6 frames (ablation) | Gemini 1.5 Flash · 4D-Bench 4D object QA | accuracy (%) | Hurt |
| increase views from 1 to 6 (ablation) | Qwen2-VL 7B · 4D-Bench 4D object captioning | GPT-Eval: 2.79 -> 2.98 | Improved |
| increase frames per view from 1 to 3 (ablation) | Qwen2-VL 7B · 4D-Bench 4D object captioning | GPT-Eval: 2.48 -> 2.89 | Improved |
| increase frames per view to 6 (ablation) | Qwen2-VL 7B · 4D-Bench 4D object captioning | GPT-Eval: 2.96 | Improved |
| increase frames per view from 6 to 9 (ablation) | Qwen2-VL 7B · 4D-Bench 4D object captioning | GPT-Eval | No clear effect |
| order frames by time instead of viewpoint (ablation) | all 14 MLLMs average · 4D-Bench 4D object QA | overall accuracy (%): 50.69 -> 50.13 | No clear effect |
| order frames by time instead of viewpoint (ablation) | Qwen2-VL 7B · 4D-Bench 4D object QA | overall accuracy (%): 56.99 -> 49.80 | Hurt |
| order frames by time instead of viewpoint (ablation) | InternVL2 8B · 4D-Bench 4D object QA | overall accuracy (%): 32.09 -> 38.88 | Improved |
| add timestamps to the prompt (ablation) | all 14 MLLMs average · 4D-Bench 4D object QA | overall accuracy (%): 50.69 -> 49.95 | No clear effect |
| add timestamps to the prompt (ablation) | VideoChat2 · 4D-Bench 4D object QA | overall accuracy (%): 32.36 -> 23.04 | Hurt |
Closest papers by meaning
- 4D LangSplat: 4D Language Gaussian Splatting via Multimodal Large Language Models (CVPR 2025 · similarity 0.91)
- Language-Image Models with 3D Understanding (ICLR 2025 · similarity 0.90)
- 4DP-QA: Scalable QA for 4D Perception in Vision Language Models (CVPR 2026 · similarity 0.90)
- Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering (ICLR 2025 · similarity 0.90)
- GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models (ICLR 2026 · similarity 0.89)
- JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.89)
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.
