Research map / Video / Video-language understanding
4D-RGPT: Toward Region-level 4D Understanding via Perceptual Distillation
Chiao-An Yang, Ryo Hachiuma, Sifei Liu, Subhashree Radhakrishnan, Raymond A. Yeh, Yu-Chiang Frank Wang, Min-Hung Chen
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 |
|---|---|---|---|
| apply P4D perceptual distillation with TPE | NVILA-Lite-8B · VSTI-Bench | accuracy: 45.2 -> 59.1 | Improved |
| apply P4D perceptual distillation with TPE | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 37.9 -> 42.2 | Improved |
| use P4D instead of direct SFT (4D-SFT) (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 40.1 -> 42.2 | Improved |
| use P4D instead of concatenating F4D features (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 39.5 -> 42.2 | Improved |
| inject F4D as positional encodings (4D-PE) (ablation) | NVILA-Lite-8B · STI-Bench | accuracy: 33.8 -> 31.3 | Hurt |
| combine latent and explicit distillation vs latent only (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 40.2 -> 42.2 | Improved |
| add depth and flow explicit distillation to latent distillation (ablation) | NVILA-Lite-8B · STI-Bench | accuracy: 34.2 -> 36.2 | Improved |
| add motion and camray explicit signals (ablation) | NVILA-Lite-8B · R4D-Bench (dynamic) | accuracy: 45.3 -> 45.4 (dynamic) | No clear effect |
| add timestamp positional encoding (TPE) to P4D (ablation) | NVILA-Lite-8B · STI-Bench | accuracy: 34.8 -> 37.6 | Improved |
| use time marks or prompts instead of TPE (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 41.0 -> 41.1 (marks) | No clear effect |
| tune vision encoder only (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 37.9 -> 35.8 | Hurt |
| tune projector and LLM fully instead of LLM LoRA only (ablation) | NVILA-Lite-8B · R4D-Bench (avg) | accuracy: 41.1 -> 42.2 | Improved |
Closest papers by meaning
- 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)
- MLLM-4D: Towards Visual-based Spatial-Temporal Intelligence (ICML 2026 · similarity 0.88)
- Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy (ICLR 2026 · similarity 0.88)
- RTV-Bench: Benchmarking MLLM Continuous Perception, Understanding and Reasoning through Real-Time Video (NeurIPS 2025 · similarity 0.87)
- 4DP-QA: Scalable QA for 4D Perception in Vision Language Models (CVPR 2026 · similarity 0.87)
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.
