Videos have change into an more and more necessary a part of our each day lives, spanning fields comparable to leisure, schooling, and communication. Understanding the content material of movies, nevertheless, is a difficult activity as movies usually comprise a number of occasions occurring at totally different time scales. For instance, a video of a musher hitching up canines to a canine sled earlier than all of them race away entails an extended occasion (the canines pulling the sled) and a brief occasion (the canines being hitched to the sled). One option to spur analysis in video understanding is through the duty of dense video captioning, which consists of temporally localizing and describing all occasions in a minutes-long video. This differs from single picture captioning and normal video captioning, which consists of describing brief movies with a single sentence.
Dense video captioning programs have extensive purposes, comparable to making movies accessible to folks with visible or auditory impairments, robotically producing chapters for movies, or enhancing the search of video moments in giant databases. Current dense video captioning approaches, nevertheless, have a number of limitations — for instance, they usually comprise extremely specialised task-specific parts, which make it difficult to combine them into highly effective basis fashions. Furthermore, they’re usually educated solely on manually annotated datasets, that are very tough to acquire and therefore are usually not a scalable resolution.
In this submit, we introduce “Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning”, to seem at CVPR 2023. The Vid2Seq structure augments a language mannequin with particular time tokens, permitting it to seamlessly predict occasion boundaries and textual descriptions in the identical output sequence. In order to pre-train this unified mannequin, we leverage unlabeled narrated movies by reformulating sentence boundaries of transcribed speech as pseudo-event boundaries, and utilizing the transcribed speech sentences as pseudo-event captions. The ensuing Vid2Seq mannequin pre-trained on hundreds of thousands of narrated movies improves the state-of-the-art on a wide range of dense video captioning benchmarks together with YouCook2, ViTT and ActivityNet Captions. Vid2Seq additionally generalizes properly to the few-shot dense video captioning setting, the video paragraph captioning activity, and the usual video captioning activity. Finally, now we have additionally launched the code for Vid2Seq right here.
 |
| Vid2Seq is a visible language mannequin that predicts dense occasion captions along with their temporal grounding in a video by producing a single sequence of tokens. |
A visible language mannequin for dense video captioning
Multimodal transformer architectures have improved the state-of-the-art on a variety of video duties, comparable to motion recognition. However it’s not easy to adapt such an structure to the complicated activity of collectively localizing and captioning occasions in minutes-long movies.
For a basic overview of how we obtain this, we increase a visible language mannequin with particular time tokens (like textual content tokens) that signify discretized timestamps within the video, just like Pix2Seq within the spatial area. Given visible inputs, the ensuing Vid2Seq mannequin can each take as enter and generate sequences of textual content and time tokens. First, this permits the Vid2Seq mannequin to grasp the temporal info of the transcribed speech enter, which is solid as a single sequence of tokens. Second, this permits Vid2Seq to collectively predict dense occasion captions and temporally floor them within the video whereas producing a single sequence of tokens.
The Vid2Seq structure features a visible encoder and a textual content encoder, which encode the video frames and the transcribed speech enter, respectively. The ensuing encodings are then forwarded to a textual content decoder, which autoregressively predicts the output sequence of dense occasion captions along with their temporal localization within the video. The structure is initialized with a highly effective visible spine and a sturdy language mannequin.
 |
| Vid2Seq mannequin overview: We formulate dense occasion captioning as a sequence-to-sequence drawback, utilizing particular time tokens to permit the mannequin to seamlessly perceive and generate sequences of tokens containing each textual semantic info and temporal localization info grounding every textual content sentence within the video. |
Large-scale pre-training on untrimmed narrated movies
Due to the dense nature of the duty, the guide assortment of annotations for dense video captioning is especially costly. Hence we pre-train the Vid2Seq mannequin utilizing unlabeled narrated movies, that are simply out there at scale. In specific, we use the YT-Temporal-1B dataset, which incorporates 18 million narrated movies protecting a variety of domains.
We use transcribed speech sentences and their corresponding timestamps as supervision, that are solid as a single sequence of tokens. We pre-train Vid2Seq with a generative goal that teaches the decoder to foretell the transcribed speech sequence given visible inputs solely, and a denoising goal that encourages multimodal studying by requiring the mannequin to foretell masked tokens given a loud transcribed speech sequence and visible inputs. In specific, noise is added to the speech sequence by randomly masking out spans of tokens.
 |
| Vid2Seq is pre-trained on unlabeled narrated movies with a generative goal (high) and a denoising goal (backside). |
Results on downstream dense video captioning benchmarks
The ensuing pre-trained Vid2Seq mannequin could be fine-tuned on downstream duties with a easy most probability goal utilizing trainer forcing (i.e., predicting the subsequent token given earlier ground-truth tokens). After fine-tuning, Vid2Seq notably improves the state-of-the-art on three normal downstream dense video captioning benchmarks (ActivityNet Captions, YouCook2 and ViTT) and two video clip captioning benchmarks (MSR-VTT, MSVD). In our paper we offer extra ablation research, qualitative outcomes, in addition to ends in the few-shot settings and within the video paragraph captioning activity.
 |
| Comparison to state-of-the-art strategies for dense video captioning (left) and for video clip captioning (proper), on the CIDEr metric (larger is best). |
Conclusion
We introduce Vid2Seq, a novel visible language mannequin for dense video captioning that merely predicts all occasion boundaries and captions as a single sequence of tokens. Vid2Seq could be successfully pretrained on unlabeled narrated movies at scale, and achieves state-of-the-art outcomes on varied downstream dense video captioning benchmarks. Learn extra from the paper and seize the code right here.
Acknowledgements
This analysis was performed by Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic and Cordelia Schmid.