Multilevel Language and Vision Integration for Text-to-Clip Retrieval

Abstract

We address the problem of text-based activity retrieval in video. Given a sentence describing an activity, our task is to retrieve matching clips from an untrimmed video. To capture the inherent structures present in both text and video, we introduce a multilevel model that integrates vision and language features earlier and more tightly than prior work. First, we inject text features early on when generating clip proposals, to help eliminate unlikely clips and thus speed up processing and boost performance. Second, to learn a fine-grained similarity metric for retrieval, we use visual features to modulate the processing of query sentences at the word level in a recurrent neural network. A multi-task loss is also employed by adding query re-generation as an auxiliary task. Our approach significantly outperforms prior work on two challenging benchmarks: Charades-STA and ActivityNet Captions.

Language Features Matter: Effective Language Representations for Vision-Language Tasks

“…Also, models that more carefully consider the effect of each word in a caption may benefit more from our improved features (e.g. [41,60] these vision-language tasks. Visual Word2Vec performs comparably amongst results for generation tasks (i.e.…”

Cross-Modal Interaction Networks for Query-Based Moment Retrieval in Videos

“…Early works study this task in constrained settings, including the fixed spatial prepositions [21,38], instruction videos [1,31,35] and ordering constraint [4,37]. Recently, unconstrained query-based moment retrieval has attracted a lot of attention [6,10,13,14,22,23,42]. These methods are mainly based on a sliding window framework, which first samples candidate moments and then ranks these moments.…”

“…That is, each frame is not only relevant to adjacent frames, but also associated with distant ones. Existing approaches often apply RNN-based temporal modeling [6], or propose R-C3D networks to learn spatiotemporal representations from raw video streams [42]. Although these methods are able to absorb contextual information for each frame, they still fail to build direct interactions between distant frames.…”

“…Early approaches [10,13,14] ignore this factor and only simply combine the query and moment features for correlation estimations. Although recent methods [6,22,23,42] have developed a cross-modal interaction by widely-used attention mechanism, they still remain in the rough one-stage interaction, for example, highlighting the crucial context information of moments by the guidance of queries [22]. Different from previous works, we adopt a multi-stage cross-modal interaction method to further exploit the potential relation of video and query contents.…”

Adversarial Representation Learning for Text-to-Image Matching

“…Text-Image Matching: Learning cross-modal embeddings has numerous applications [61,69] ranging from PINs using facial and voice information [37], to generative feature learning [15] and domain adaptation [63,65]. Nagrani et al [37] demonstrated that a joint representation can be learned from facial and voice information and introduced a curriculum learning strategy [3,45,46] to perform hard negative mining during training.…”