CONQUER_RVMR
This repository contains the XML model for the baseline of the Ranked Video Moment Retrieval (RVMR) task. The associated paper is titled "Video Moment Retrieval in Practical Setting: A Dataset of Ranked Moments for Imprecise Queries."
The main repository of the paper is TVR-Ranking, and this model is adapted from CONQUER. The environment setup is the same as for RelocNet_RVMR, as detailed in the TVR-Ranking repository.
CONQUER leverages video retrieval results from HERO. We continue to use these results when training on our TVR-Ranking dataset. Note that, because the HERO results are obtained from the TVR dataset, there could be a data leak issue in our task setting. However, this issue is negligible for two reasons: (i) the queries used in our setting is imprecise query with query re-written, and (ii) a query has multiple ground truth moments in our task setting, which was not annotated in the original TVR dataset.
Performance
| Model | Train Set Top N | IoU=0.3 | IoU=0.5 | IoU=0.7 | |||
|---|---|---|---|---|---|---|---|
| Val | Test | Val | Test | Val | Test | ||
| NDCG@10 | |||||||
| CONQUER | 1 | 0.0999 | 0.0859 | 0.0844 | 0.0709 | 0.0530 | 0.0512 |
| CONQUER | 20 | 0.2406 | 0.2249 | 0.2222 | 0.2104 | 0.1672 | 0.1517 |
| CONQUER | 40 | 0.2450 | 0.2219 | 0.2262 | 0.2085 | 0.1670 | 0.1515 |
| NDCG@20 | |||||||
| CONQUER | 1 | 0.0952 | 0.0835 | 0.0808 | 0.0687 | 0.0526 | 0.0484 |
| CONQUER | 20 | 0.2130 | 0.1995 | 0.1976 | 0.1867 | 0.1527 | 0.1368 |
| CONQUER | 40 | 0.2183 | 0.1968 | 0.2022 | 0.1851 | 0.1524 | 0.1365 |
| NDCG@40 | |||||||
| CONQUER | 1 | 0.0974 | 0.0866 | 0.0832 | 0.0718 | 0.0557 | 0.0510 |
| CONQUER | 20 | 0.2029 | 0.1906 | 0.1891 | 0.1788 | 0.1476 | 0.1326 |
| CONQUER | 40 | 0.2080 | 0.1885 | 0.1934 | 0.1775 | 0.1473 | 0.1323 |
Quick Start
Modify the path in run_disjoint_top20.sh and then execute the script:
sh run_disjoint_top20.sh
Feel free to contribute or raise issues for any problems encountered.