Abstract
With the explosive growth of video data in real applications, near-duplicate video retrieval (NDVR) has become indispensable and challenging, especially for short videos. However, all existing NDVR datasets are introduced for long videos. Furthermore, most of them are small-scale and lack of diversity due to the high cost of collecting and labeling near-duplicate videos. In this paper, we introduce a large-scale short video dataset, called SVD, for the NDVR task. SVD contains over 500,000 short videos and over 30,000 labeled videos of near-duplicates. We use multiple video mining techniques to construct positive/negative pairs. Furthermore, we design temporal and spatial transformations to mimic user-attack behavior in real applications for constructing more difficult variants of SVD. Experiments show that existing state-of-the-art NDVR methods, including real-value based and hashing based methods, fail to achieve satisfactory performance on this challenging dataset.
Introduction
Near-Duplicate Video Retrieval (NDVR) is the task of identifying videos that are visually almost identical to a query video but have been slightly modified—often to bypass copyright filters or for creative re-purposing. While NDVR is a mature field, existing benchmarks like CCWEB or VCDB focus almost exclusively on long-form videos.
The explosion of platforms like TikTok and Instagram has shifted the landscape toward amateur, short-form content. These videos present unique challenges: they are often filmed on mobile devices, edited with aggressive filters, or reformatted to fit different screen orientations. SVD (Short Video Dataset) is the first large-scale benchmark designed to address these specific "wild" short-form data challenges.
The SVD Dataset
SVD consists of 1,206 query videos, 34,020 labeled videos (positives and negatives), and a massive distraction set of 526,787 unlabeled videos. The defining characteristic of SVD is its focus on short durations, with an average video length of just 17.33 seconds. This stands in stark contrast to legacy datasets where durations often exceed several minutes.
As shown in the distribution analysis, SVD captures the high-frequency, short-duration nature of modern social media content, whereas CCWEB and VCDB are skewed toward much longer professional or semi-professional productions.
Constructing a Challenging Benchmark
To ensure the dataset is robust and difficult for current algorithms, the researchers employed a multi-stage pipeline:
- Iterative Mining: Initial candidates were found using basic visual features and human-labeled. These positives were then used as new queries to find even more subtle duplicates.
- Hard Negative Mining: The dataset specifically includes "hard negatives"—videos that look visually similar (e.g., same setting or similar-looking person) but contain different content.
- Simulated User Attacks: To mimic real-world behavior, the authors applied temporal transformations (speed changes) and spatial transformations (cropping, black borders, and rotation).
Key Insights and Experimental Results
The researchers tested several state-of-the-art "Real-Value" methods (like Deep Metric Learning) and "Hashing" methods. The results revealed a significant performance gap when moving from long-form to short-form content:
- Performance Degradation: The best-performing model achieved a Mean Average Precision (MAP) of 97.01% on the older CCWEB dataset but dropped to 78.47% on SVD.
- The Rotation Weakness: Standard convolutional neural network (CNN) features proved highly sensitive to orientation. On a rotated version of SVD, the accuracy of top-tier models cratered to approximately 15.59%.
- Information Density: Because short videos have less temporal information, small edits (like adding a caption or a border) have a disproportionately large impact on the video's digital signature compared to longer videos where such noise can be "averaged out."
Conclusion
The SVD dataset highlights a major gap in current NDVR systems: they are not robust against the spatial transformations common in mobile-centric short videos. By providing a large-scale, diverse "playground," SVD enables the community to develop the next generation of rotation-invariant and transformation-robust features necessary for modern copyright and recommendation systems.
Summary updated