X-Pool: Cross-Modal Language-Video Attention for Text-Video Retrieval
In text-video retrieval, the objective is to learn a cross-modal similarity function between a text and a video that ranks relevant text-video pairs higher than irrelevant pairs. However, videos inherently express a much wider gamut of information than texts. Instead, texts often capture sub-regions of entire videos and are most semantically similar to certain frames within videos. Therefore, for a given text, a retrieval model should focus on the text's most semantically similar video sub-regions to make a more relevant comparison. Yet, most existing works aggregate entire videos without directly considering text. Common text-agnostic aggregations schemes include mean-pooling or self-attention over the frames, but these are likely to encode misleading visual information not described in the given text. To address this, we propose a cross-modal attention model called X-Pool that reasons between a text and the frames of a video. Our core mechanism is a scaled dot product attention for a text to attend to its most semantically similar frames. We then generate an aggregated video representation conditioned on the text's attention weights over the frames. We evaluate our method on three benchmark datasets of MSR-VTT, MSVD and LSMDC, achieving new state-of-the-art results by up to 12% in relative improvement in Recall@1. Our findings thereby highlight the importance of joint text-video reasoning to extract important visual cues according to text. Full code and demo can be found at: https://layer6ai-labs.github.io/xpool/
PDF Abstract CVPR 2022 PDF CVPR 2022 AbstractCode
Datasets
Visual Question Answering
MSR-VTT
MSVD
LSMDC
Results from the Paper
Ranked #17 on
Video Retrieval
on LSMDC
(using extra training data)
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Uses Extra Training Data |
Benchmark |
|---|---|---|---|---|---|---|---|
| Video Retrieval | LSMDC | X-Pool | text-to-video R@1 | 25.2 | # 17 | ||
| text-to-video R@5 | 43.7 | # 15 | |||||
| text-to-video R@10 | 53.5 | # 17 | |||||
| text-to-video Median Rank | 8.0 | # 6 | |||||
| video-to-text R@1 | 22.7 | # 11 | |||||
| video-to-text R@5 | 42.6 | # 8 | |||||
| video-to-text R@10 | 51.2 | # 8 | |||||
| video-to-text Median Rank | 10.0 | # 5 | |||||
| text-to-video Mean Rank | 53.2 | # 6 | |||||
| video-to-text Mean Rank | 47.4 | # 6 | |||||
| Video Retrieval | MSR-VTT-1kA | X-Pool | text-to-video Mean Rank | 14.3 | # 17 | ||
| text-to-video R@1 | 46.9 | # 28 | |||||
| text-to-video R@5 | 72.8 | # 27 | |||||
| text-to-video R@10 | 82.2 | # 29 | |||||
| text-to-video Median Rank | 2 | # 10 | |||||
| video-to-text R@1 | 44.4 | # 19 | |||||
| video-to-text R@5 | 73.3 | # 17 | |||||
| video-to-text R@10 | 84.0 | # 13 | |||||
| video-to-text Median Rank | 2.0 | # 7 | |||||
| video-to-text Mean Rank | 9.0 | # 13 | |||||
| Video Retrieval | MSVD | X-Pool | text-to-video R@1 | 47.2 | # 17 | ||
| text-to-video R@5 | 77.4 | # 13 | |||||
| text-to-video R@10 | 86.0 | # 12 | |||||
| text-to-video Median Rank | 2.0 | # 8 | |||||
| text-to-video Mean Rank | 9.3 | # 9 | |||||
| video-to-text R@1 | 66.4 | # 11 | |||||
| video-to-text R@5 | 90.0 | # 8 | |||||
| video-to-text R@10 | 94.2 | # 8 | |||||
| video-to-text Median Rank | 1.0 | # 1 | |||||
| video-to-text Mean Rank | 3.3 | # 5 |
