DIBS: Enhancing Dense Video Captioning with Unlabeled Videos via Pseudo Boundary Enrichment and Online Refinement

TL;DR

DIBS tackles dense video captioning by exploiting unlabeled videos to generate high-quality event captions and pseudo boundaries with the help of diverse LLMs. It introduces a caption-aware boundary generation method with soft time constraints and an online boundary refinement strategy that iteratively improves pseudo targets during training, integrated with PDVC-style pretraining. Empirical results on YouCook2 and ActivityNet show substantial gains in caption quality and competitive or superior boundary localization, often outperforming Vid2Seq while using far less unlabeled data. The approach demonstrates a practical pathway to leverage large-scale unlabeled video data for DVC, reducing annotation costs and enhancing cross-domain generalization, especially when the pretraining data aligns with the target domain.

Abstract

We present Dive Into the BoundarieS (DIBS), a novel pretraining framework for dense video captioning (DVC), that elaborates on improving the quality of the generated event captions and their associated pseudo event boundaries from unlabeled videos. By leveraging the capabilities of diverse large language models (LLMs), we generate rich DVC-oriented caption candidates and optimize the corresponding pseudo boundaries under several meticulously designed objectives, considering diversity, event-centricity, temporal ordering, and coherence. Moreover, we further introduce a novel online boundary refinement strategy that iteratively improves the quality of pseudo boundaries during training. Comprehensive experiments have been conducted to examine the effectiveness of the proposed technique components. By leveraging a substantial amount of unlabeled video data, such as HowTo100M, we achieve a remarkable advancement on standard DVC datasets like YouCook2 and ActivityNet. We outperform the previous state-of-the-art Vid2Seq across a majority of metrics, achieving this with just 0.4% of the unlabeled video data used for pre-training by Vid2Seq.

Figures (6)

• Figure 1: Comparison between Vid2Seq Vid2Seq2023 and ours. (a) Raw subtitles extracted from videos. (b) Vid2Seq Vid2Seq2023 directly converts the raw text and timestamps of subtitles into pseudo event boundaries and captions for pretraining. (c) Our proposed framework utilizes LLMs to generate rich and accurate captions for events from raw narrations. Subsequently, corresponding pseudo boundaries can be generated using these captions through our devised optimization algorithm (cf. Sec. \ref{['sec:pseudo_box']}).
• Figure 2: Prior range division and frame processing in pseudo boundary generation. Frames showcasing top-$k$ global similarity to Event-3 are indicated by deep green points, contrasting with lighter green points denoting similarity outside the top-$k$ range. The blue box delineates the evenly divided prior range, while the orange box signifies the adjusted range after an iteration. The pseudo boundary generation process is outlined in \ref{['alg:pseudo_box_generation']}.
• Figure 3: The process of online pseudo boundary refinement.
• Figure 4: Word cloud of predicted captions from our DIBS w/ and w/o pretraining. A larger word indicates a higher frequency of its occurrence in the predicted captions.
• Figure 5: Visualization of qualitative results of our DIBS model. Video examples are selected from the validation set of YouCook2.