Research map / Vision-language and multimodal / CLIP and zero-shot
InstAP: Instance-Aware Vision-Language Pre-Train for Spatial-Temporal Understanding
Ashutosh Kumar, Rajat Saini, Jingjing Pan, Mustafa Erdogan, Mingfang Zhang, Betty Le Dem, Norimasa Kobori, Quan Kong
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 instance-aware loss L_inst (ablation) | InstAP · InstVL-1K (video) instance | mean recall: 57.71 -> 75.32 | Improved |
| add instance-aware loss L_inst (ablation) | InstAP · InstVL-1K (img-zero) instance | mean recall: 49.98 -> 63.94 | Improved |
| add instance-aware loss L_inst (ablation) | InstAP · InstVL-1K (video) global | mean recall: 91.55 -> 97.03 | Improved |
| add instance-aware loss L_inst (ablation) | InstAP · DiDeMo | mean recall: 65.98 -> 70.01 | Improved |
| add learnable instance temperature (ablation) | InstAP · InstVL-1K (img) | mean recall: 59.10 -> 67.19 | Improved |
| add weighted instance loss (lambda 0.1) (ablation) | InstAP · InstVL-1K (video) | mean recall: 55.22 -> 58.16 | Improved |
| add caption sub-sampling (ablation) | InstAP · InstVL-1K (img) | mean recall: 68.17 -> 71.65 | Improved |
| add 50K instance video trajectory data (ablation) | InstAP · InstVL-1K (video) | mean recall: 58.97 -> 75.32 | Improved |
| treat instance captions as global captions (g+i) vs instance-aware training | UMT-L baseline · InstVL-10K (img) | T2V R@1: 44.05 -> 34.83 | Hurt |
| fine-tune UMT-L on InstVL global captions only | UMT-L · MSR-VTT | R@1: 39.7 -> 35.4 | Hurt |
| use InstAP instead of UMT-L | zero-shot retrieval · DiDeMo | R@1: 47.0 -> 54.0 | Improved |
| use InstAP instead of UMT-L | zero-shot retrieval · LSMDC | R@1: 26.0 -> 23.5 | Hurt |
Closest papers by meaning
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · similarity 0.88)
- INST-IT: Boosting Instance Understanding via Explicit Visual Prompt Instruction Tuning (NeurIPS 2025 · similarity 0.88)
- From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection (ICML 2025 · similarity 0.88)
- Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding (NeurIPS 2026 · similarity 0.88)
- RegionGPT: Towards Region Understanding Vision Language Model (CVPR 2024 · similarity 0.87)
- Object-centric binding in Contrastive Language-Image Pretraining (NeurIPS 2025 · 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.
