Research map / Point clouds / Point cloud classification
CL4D: Contrastive Language–4D Pretraining for Vision-Language Reasoning in Dynamic Scenes
Kumal Hewagamage, Isuranga Senavirathne, Sasika Amarasinghe, Hasitha Gallella, Dulanga Weerakoon, Vigneshwaran Subbaraju, Ranga Rodrigo
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 CL4D contrastive 4D encoder instead of P4Transformer | contrastive text-motion alignment · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 53.57 -> 70.32 | Improved |
| use CL4D instead of Motion PointNet | contrastive text-motion alignment · DynAction4D ObjInteractions | Text-to-motion R@1 (Batch): 41.37 -> 49.40 | Improved |
| use CL4D instead of Motion PointNet | contrastive text-motion alignment · DynAction4D Cluttered | Text-to-motion R@1 (Batch): 41.71 -> 55.07 | Improved |
| train the text encoder jointly instead of freezing it (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 56.63 -> 68.43 | Improved |
| use InfoNCE loss instead of SigLIP loss (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 66.97 -> 68.43 | Improved |
| initialize temporal encoder from pretrained 2D ViT (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 63.32 -> 68.43 | Improved |
| keep the correct temporal frame order instead of shuffling (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 60.97 -> 68.43 | Improved |
| change ViT backbone size (Tiny, Small, Base, Large) (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 68.79 / 69.61 / 68.43 / 66.95 | Mixed |
| use CL4D-mini (reduced spatial encoder, ViT-Tiny) (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 68.43 -> 67.18 | No clear effect |
| apply partial point clouds or random reconstruction error (ablation) | CL4D-mini · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 67.18 -> 66.81 (partial), 65.85 (random error) | No clear effect |
| increase training batch size from 16 to 72 (ablation) | CL4D (InfoNCE) · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 58.71 -> 70.32 | Improved |
| increase number of input frames from 4 to 32 (ablation) | CL4D · DynAction4D HumanOnly | Text-to-motion R@1 (Batch): 49.85 -> 65.62 | Improved |
Closest papers by meaning
- 4DP-QA: Scalable QA for 4D Perception in Vision Language Models (CVPR 2026 · similarity 0.90)
- Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding (NeurIPS 2025 · similarity 0.90)
- Feature4X: Bridging Any Monocular Video to 4D Agentic AI with Versatile Gaussian Feature Fields (CVPR 2025 · similarity 0.89)
- 4D-VLA: Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration (NeurIPS 2025 · similarity 0.89)
- Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering (ICLR 2025 · similarity 0.89)
- 4D LangSplat: 4D Language Gaussian Splatting via Multimodal Large Language Models (CVPR 2025 · 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.
