Research map / Video / Video-language understanding
Text-video retrieval
Retrieves videos from text using coarse-to-fine, probabilistic and ambiguity-aware representations.
59 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Video-language understanding, keyword cluster “retrieval · text · moment”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 3 of 4 tracked benchmarks.
13 papers in 2023–24 and 34 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.8% to 4.7%.
Open questions the data points to
- ActivityNet Captions (SumR): best went from 146 (up to 2024) to 202.1 in 2025 — Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · 0 citations)
- TVR (SumR): best went from 176.6 (up to 2024) to 263.1 in 2025 — Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · 0 citations)
- VATEX, text-to-video (R@1): best went from 66.6 (up to 2024) to 68.4 in 2025 — Narrating the Video: Boosting Text-Video Retrieval via Comprehensive Utilization of Frame-Level Captions (CVPR 2025 · 11 citations)
- Papers report MS-SL on TVR (SumR) anywhere from 172.4 to 251: comparisons against it depend on whose number you take.
What looks settled
- DiDeMo text-to-video (R@1): best up to 2024 53.3, best in 2025–26 53.4 from 4 papers, at most 0.2% better, if at all.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Collaborative Summarization of Topic-Related Videos (CVPR 2017 · 81 citations)
- 2019 · Multilevel Language and Vision Integration for Text-to-Clip Retrieval (AAAI 2019 · 340 citations)
- 2020 · A Proposal-Based Approach for Activity Image-to-Video Retrieval (AAAI 2020 · 18 citations)
- 2021 · On Semantic Similarity in Video Retrieval (CVPR 2021 · 1 citation)
- 2022 · X-Pool: Cross-Modal Language-Video Attention for Text-Video Retrieval (CVPR 2022 · 223 citations)
- 2023 · Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval? (CVPR 2023 · 138 citations)
- 2024 · Text Is MASS: Modeling as Stochastic Embedding for Text-Video Retrieval (CVPR 2024 · 77 citations)
- 2025 · Learning Audio-guided Video Representation with Gated Attention for Video-Text Retrieval (CVPR 2025 · 12 citations)
- 2026 · ReTrack: Evidence-Driven Dual-Stream Directional Anchor Calibration Network for Composed Video Retrieval (AAAI 2026 · 7 citations)
Earliest papers here
- Collaborative Summarization of Topic-Related Videos (CVPR 2017 · 81 citations)
- Multilevel Language and Vision Integration for Text-to-Clip Retrieval (AAAI 2019 · 340 citations)
- Dual Encoding for Zero-Example Video Retrieval (CVPR 2019 · 301 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
DiDeMo text-to-video · R@1 (higher is better)
6 papers, 2023–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 52 | Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval? (CVPR 2023 · 138 citations) |
| 2024 | 53.3 | Text Is MASS: Modeling as Stochastic Embedding for Text-Video Retrieval (CVPR 2024 · 77 citations) |
| 2025 | 53.4 | Narrating the Video: Boosting Text-Video Retrieval via Comprehensive Utilization of Frame-Level Captions (CVPR 2025 · 11 citations) |
ActivityNet Captions · SumR (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 146 | GMMFormer: Gaussian-Mixture-Model Based Transformer for Efficient Partially Relevant Video Retrieval (AAAI 2024 · 17 citations) |
| 2025 | 202.1 | Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · 0 citations) |
TVR · SumR (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 176.6 | GMMFormer: Gaussian-Mixture-Model Based Transformer for Efficient Partially Relevant Video Retrieval (AAAI 2024 · 17 citations) |
| 2025 | 263.1 | Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · 0 citations) |
VATEX, text-to-video · R@1 (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 66.6 | Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval? (CVPR 2023 · 138 citations) |
| 2025 | 68.4 | Narrating the Video: Boosting Text-Video Retrieval via Comprehensive Utilization of Frame-Level Captions (CVPR 2025 · 11 citations) |
What has been tried
41 papers here report 489 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, no clear effect 4%, hurt 17%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add a similarity-aware radius module | MSRVTT text-to-video · R@1 | 46.9 -> 49.1 | Helped | Text Is MASS: Modeling as Stochastic Embedding for Text-Video Retrieval (CVPR 2024 · 77 citations) |
| Combine visual and text query modalities | WebVid-CoVR-Test · R@1 | 38.89 -> 53.13 | Helped | CoVR: Learning Composed Video Retrieval from Web Video Captions (AAAI 2024 · 50 citations) |
| add AFA to CLIP4Clip | MSRVTT-1k · SumR | 195.5 -> 196.3 | Helped | Holistic Features are almost Sufficient for Text-to-Video Retrieval (CVPR 2024 · 28 citations) |
| Use original ASR transcripts instead of refined captions | Ego4d cooking validation · CIDER | 166.5 vs. 173.8 | Hurt | Retrieval-Augmented Egocentric Video Captioning (CVPR 2024 · 24 citations) |
| Add temporal modeling with TRM to the Down&Up adapter | MSR-VTT, T2V · R@Sum | 192.0 -> 197.5 | Helped | MV-Adapter: Multimodal Video Transfer Learning for Video Text Retrieval (CVPR 2024 · 22 citations) |
| add the text-patch interaction feature | MSRVTT-1K-Test t2v · Mean | 67.5 -> 67.8 | Helped | Towards Efficient and Effective Text-to-Video Retrieval with Coarse-to-Fine Visual Representation Learning (AAAI 2024 · 18 citations) |
| Use GMMFormer blocks instead of vanilla Transformer blocks | TVR · SumR | 163.5 -> 172.9 | Helped | GMMFormer: Gaussian-Mixture-Model Based Transformer for Efficient Partially Relevant Video Retrieval (AAAI 2024 · 17 citations) |
| Use video, detailed description, and change text together | WebVid-CoVR test set · R@1 | 56.26 -> 60.12 | Helped | Composed Video Retrieval via Enriched Context and Discriminative Embeddings (CVPR 2024 · 16 citations) |
| Add audio using a simplified fusion transformer | MSR-VTT 9k, text-to-video · R@1 | 45.4 -> 47.1 | Helped | Learning Audio-guided Video Representation with Gated Attention for Video-Text Retrieval (CVPR 2025 · 12 citations) |
| Add narration matching | MSR-VTT 1k-A · R@1 | 44.5 -> 48.6 | Helped | Narrating the Video: Boosting Text-Video Retrieval via Comprehensive Utilization of Frame-Level Captions (CVPR 2025 · 11 citations) |
| combine frame-level and video-level MMS scores instead of using frame-level MMS alone | MSR-VTT · R@1 | 44.3 -> 48.1 | Helped | Video-ColBERT: Contextualized Late Interaction for Text-to-Video Retrieval (CVPR 2025 · 11 citations) |
| Replace Mamba with Transformer | MSR-VTT · R@1 | 44.8 -> 43.0 | Hurt | MUSE: Mamba Is Efficient Multi-scale Learner for Text-video Retrieval (AAAI 2025 · 9 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
What reviewers will expect
Datasets and baselines named in the 20 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MSR-VTT · 11 of 20 papers
- DiDeMo · 9 of 20 papers
- ActivityNet Captions · 6 of 20 papers
- MSVD · 6 of 20 papers
- LSMDC · 5 of 20 papers
- ActivityNet · 5 of 20 papers
- VATEX · 4 of 20 papers
- TVR · 3 of 20 papers
Baselines
- CLIP4Clip · 10 of 20 papers
- X-Pool · 8 of 20 papers
- UATVR · 6 of 20 papers
- TS2-Net · 5 of 20 papers
- T-MASS · 4 of 20 papers
- Cap4Video · 4 of 20 papers
- MS-SL · 3 of 20 papers
- DL-DKD · 3 of 20 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 47 papers; from OpenAlex and the papers' affiliations).
- Sungkyunkwan University · 4 papers
- Hefei University of Technology · 3 papers
- Amazon · 3 papers
- Peng Cheng Laboratory · 2 papers
- Sichuan University · 2 papers
- Nanjing University · 2 papers
- Rochester Institute of Technology · 2 papers
- DEVCOM Army Research Laboratory · 2 papers
Use it
Download BibTeX of 32 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Video-language pretraining (82 papers)
- Video moment retrieval (41 papers)
Elsewhere in the topic
- Long video understanding (236 papers)
- Video LLMs with fine-grained understanding (229 papers)
- Event and anomaly reasoning in video (184 papers)
- Temporal video grounding (88 papers)
- Dense video captioning (65 papers)
- Video question answering (64 papers)
- Encoder-decoder video captioning (52 papers)
- Egocentric video understanding (46 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).