Research map / Vision-language and multimodal / CLIP and zero-shot
MobileViCLIP: An Efficient Video-Text Model for Mobile Devices
Min Yang, Zihan Jia, Zhilin Dai, Sheng Guo, Limin Wang
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 |
|---|---|---|---|
| add spatiotemporal RepMixer to MobileCLIP (ablation) | MobileCLIP · MSR-VTT fine-tuned T2V | R@1: 38.4 -> 39.5 | Improved |
| add spatiotemporal attention without TPE (ablation) | MobileCLIP · MSR-VTT fine-tuned T2V | R@1: 38.4 -> 39.1 | Improved |
| add learnable temporal positional encodings to spatiotemporal attention (ablation) | MobileViCLIP · MSR-VTT fine-tuned T2V | R@1: 39.1 -> 39.6 | Improved |
| combine RepMixer and spatiotemporal attention with TPE (ablation) | MobileViCLIP · MSR-VTT fine-tuned T2V | R@1: 39.5 -> 40.1 | Improved |
| also train the text branch rather than freezing it (ablation) | MobileViCLIP · MSR-VTT fine-tuned T2V | R@1: 40.1 -> 39.7 | Hurt |
| fine-tune MobileCLIP-S0 on InternVid-10M-FLT (no temporal modules) | MobileCLIP-S0 · MSR-VTT zero-shot | T2V R@1: 33.2 -> 37.2 | Improved |
| add temporal modules on top of InternVid fine-tuning | MobileCLIP-S0 · MSR-VTT zero-shot | T2V R@1: 37.2 -> 38.7 | Improved |
| add temporal modules on top of InternVid fine-tuning | MobileCLIP-S2 · MSR-VTT zero-shot | T2V R@1: 41.2 -> 42.5 | Improved |
| add temporal modules on top of InternVid fine-tuning | MobileCLIP-S2 · SSV2-MC zero-shot | top-1 accuracy: 44.1 -> 46.1 | Improved |
| MobileViCLIP-Small vs InternVideo2-S14 on ActivityNet-style long video | MobileViCLIP-Small · SSV2-MC zero-shot | top-1 accuracy: 46.4 -> 46.1 | Hurt |
| use MobileViCLIP-Small features instead of MobileCLIP-S0(10M) for temporal grounding | CG-DETR · QVHighlights | mAP avg: 43.8 -> 46.3 | Improved |
| use MobileViCLIP-Small features instead of CoCa ViT-L/14 for zero-shot temporal action detection | T3AL · THUMOS14 | avg mAP: 4.8 -> 5.3 | Improved |
Closest papers by meaning
- MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training (CVPR 2024 · similarity 0.88)
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · similarity 0.85)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.85)
- MobileInst: Video Instance Segmentation on the Mobile (AAAI 2024 · similarity 0.83)
- Scalable Video-to-Dataset Generation for Cross-Platform Mobile Agents (CVPR 2025 · similarity 0.83)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.83)
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.
