Research map / Video / Video-language understanding
MomentSeeker: A Task-Oriented Benchmark For Long-Video Moment Retrieval
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MomentSeeker, a human-annotated benchmark for long-video moment retrieval with text, image, and video queries, and evaluates generation-based MLLMs and retrieval-based embedders on it.
Problem
Existing moment retrieval benchmarks use short videos with simplistic caption queries, while long-video understanding benchmarks evaluate only end-to-end answer accuracy, not whether key moments are accurately located in long videos.
Approach
A task-oriented long-video moment retrieval benchmark with ~1200-second average videos across diverse domains, nine LVU-style task types organized into global, event, and object levels, and text-only, image-conditioned, and video-conditioned queries, all with human-annotated fine-grained timestamps.
Key results
- Gemini-2.5-Pro achieves the best generation-based result with 29.6 R@1 and 31.4 mAP@5 overall, while Qwen2.5-VL-72B reaches 17.2 R@1.
- Best retriever InternVideo2 reaches 19.7 R@1 and 26.6 mAP@5 overall; InternVL3-8B reaches only 5.9 R@1.
- Most models score lower on image- and video-conditioned queries than on text-only queries, and performance declines as video duration grows.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase Qwen2.5VL-7B input frames from 96 to 768 (ablation) | Qwen2.5VL-7B · MomentSeeker overall | R@1: 4.5 to 8.1 | Helped |
| Increase InternVL3-38B input frames from 64 to 128 (ablation) | InternVL3-38B · MomentSeeker overall | R@1: 15.9, 15.8, and 15.6 | No clear effect |
| Use stricter IoU thresholds (ablation) | All evaluated models · MomentSeeker | accuracy | Hurt |
| Use a larger generation model | Generation-based methods · MomentSeeker overall | R@1: 5.9; 15.8 | Helped |
| Use generation-based methods instead of retrieval-based methods | MomentSeeker overall · MomentSeeker overall | R@1: 12.4; 5.9 | Mixed |
| Use image- or video-conditioned queries instead of text-only queries | Most evaluated models · MomentSeeker query modalities | retrieval performance | Hurt |
Surprising findings
- Retrieval-based methods with 148M-parameter MM-Ret outperform the 8B generation model InternVL3 by a large margin (12.4 vs 5.9 R@1).
- Qwen2.5-VL-72B scores only 17.2 R@1 despite 79.1 on Video-MME, indicating that answer accuracy does not imply accurate grounding.
Limitations
- Generation-based models evaluated on a random 180-sample subset for GPT-4o due to high cost.
- Generation-based models are limited by context length and uniform frame downsampling for long videos.
Experimental setup
Datasets: MomentSeeker
Benchmarks: MomentSeeker
Metrics: R@1 mAP@5
Baselines: GPT-4o Gemini-2.5-Pro TimeChat Lita Qwen2.5-VL InternVL3 E5V InternVideo2 LanguageBind CoVR UniIR MM-Ret
Builds on: Video-MME MLVU LongVideoBench V-NIAH
Closest papers by meaning
- When One Moment Isn't Enough: Multi-Moment Retrieval with Cross-Moment Interactions (NeurIPS 2025 · similarity 0.89)
- Fewer Steps, Better Performance: Efficient Cross-Modal Clip Trimming for Video Moment Retrieval Using Language (AAAI 2024 · similarity 0.89)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.89)
- Generative Video Diffusion for Unseen Novel Semantic Video Moment Retrieval (AAAI 2025 · similarity 0.88)
- Transferable Video Moment Localization by Moment-Guided Query Prompting (AAAI 2024 · similarity 0.88)
- Anchor-Aware Similarity Cohesion in Target Frames Enables Predicting Temporal Moment Boundaries in 2D (CVPR 2025 · similarity 0.88)
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.