Contrastive learning for video reuse detection

Tristan J. B. Cann; Arjun Biswas

cs.CVOct 12, 2026

Contrastive learning for video reuse detection

Authors: Tristan J. B. Cann, Arjun Biswas

Abstract

The incorporation of footage from other sources has long been a concern from a copyright perspective, as such usage can harm the interests of rights holders. Recent trends in the ``remixing'' of social media content and the development of bespoke stock libraries for specialist causes mean that methods to identify content reuse will have a range of applications. In this paper, we develop a framework based on a Siamese network fine tuned using contrastive learning to identify potential sources for key frames in a given video. This framework makes two contributions to analysis of large video datasets. Firstly, we demonstrate its utility at reducing complex video data to a set of key visuals typically less than 10% of the size of the full set. Secondly, we show that these reduced sets allow for efficient and effective identification of component footage. We test this method over a set of videos compiled from a stock footage library and demonstrate its effectiveness for constituent video identification with average accuracy greater than 0.95.

Figures & tables

Figure 1: Overall approach for the comparison of videos. We use our trained model to extract a set of representative frames from both the reference and test videos, followed by using the same model to identify similar frames between both videos.

Figure 2: Overview of our contrastive learning pipeline. Input images are augmented to generate positive (similar) pairs, while different images are paired to form negative (dissimilar) pairs. Each image is processed by a shared image encoder followed by a shared fully connected projection layer to produce embedding representations, which are then used for contrastive learning. The bridge image was taken by the authors, the parrot image is taken from ImageNet100 [ 18 ] .

Panel of Figure 3 — Figure 3: Examples of the augmentations applied to each the relevant images. The raw image was taken by the authors.

Panel of Figure 4 — Figure 4: Performance statistics for our method when applied to the synthetic videos using the DenseNet-121, ResNet-50 and Clip-ViT-Base-Patch-16 backbones. The method displays a preference for recall over precision, and lower precision is the cause for lower balanced accuracy in most cases.

Figure 5: False positive and false negative proportions for each Pixabay video in the dataset, normalised by the number of composite videos they appear in. Some videos are shown to be hard to detect (false negative rate of 1) or commonly detected incorrectly (false positive rate $>5$ ).

Panel of Figure 6 — Figure 6: Recall@( $k$ ) performance for $k=1,6,12$ . Each cell illustrates the proportion of composite videos correctly identifying the relevant number of included clips.

Figure 7: Mean proportion of all frames identified as representative frames at different thresholds over all videos (clips and composites). Shaded regions indicate one standard deviation from the mean.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Video URL	Search term
https://pixabay.com/videos/people-walking-sun-sky-traffic-197531/	people
https://pixabay.com/videos/human-person-walk-the-crowd-women-95193/	people
https://pixabay.com/videos/boys-girls-group-friends-young-51337/	people
https://pixabay.com/videos/hug-friends-talk-nature-people-159243/	people
https://pixabay.com/videos/dance-music-to-dance-club-disco-1670/	people
https://pixabay.com/videos/charity-village-people-help-ethnic-139400/	people

Appendix

Table S1 : Video URLs for “people” and “city” footage collected from Pixabay in February 2025.

Video URL	Search term
https://pixabay.com/videos/trees-forest-autumn-scenic-4k-185726/	landscape
https://pixabay.com/videos/beach-sea-ocean-sand-drone-aerial-186115/	landscape
https://pixabay.com/videos/ocean-waves-rocks-sand-coast-213026/	landscape
https://pixabay.com/videos/beach-coastline-travel-tourism-161045/	landscape
https://pixabay.com/videos/ocean-beach-sunset-waves-dawn-183960/	landscape
https://pixabay.com/videos/bird-seagull-boat-ocean-sea-223461/	landscape

Appendix

Table S2 : Video URLs for “landscape” and “climate” footage collected from Pixabay in February 2025.

Video URL	Search term
https://pixabay.com/videos/frog-reeds-green-113403/	wildlife
https://pixabay.com/videos/mountains-forest-nature-sky-clouds-168787/	wildlife
https://pixabay.com/videos/sea-ocean-seagulls-birds-sunset-140111/	wildlife
https://pixabay.com/videos/snail-mollusk-creeping-animal-169471/	wildlife
https://pixabay.com/videos/flower-bee-anemone-pink-insect-178826/	wildlife
https://pixabay.com/videos/flower-peony-bee-insect-pollinate-165065/	wildlife

Appendix

Table S3 : Video URLs for “wildlife” footage collected from Pixabay in February 2025.

Figure S1 : Number of original videos within a given similarity threshold of each composite video. The solid line averages over all composite videos, and the shaded region indicates a range of one standard deviation around the mean.

Video Representation Learning

Video representation learning develops features that capture appearance, motion, and temporal structure for recognition, prediction, and control.

See all 143 papers

MatchLM2Lite: A Scalable MLLM-to-Lite Framework for Reproduced Content Identification
Jun 10, 2026Xiaotian Fan, Hiok Hian Ong, David Yuchen Wang +3Multimodal Large Language Models
Tiktok Singapore, Singapore · Tiktok National University of Singapore School of Computing Singapore, Singapore · Tiktok Content Intelligence San Jose, CA, USA +1
MLT-Dedup: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching
Jun 10, 2026David Yuchen Wang, Haoying Li, Hailun Xu +6Video Representation Learning
TikTok Singapore, Singapore · School of Computing National University of Singapore Singapore, Singapore · TikTok San Jose, CA, USA
Copyright Leakage Detection in AI Generated Videos
Oct 20, 2026Shaif Chowdhury, Moshiur Rahman Tonmoy
Department of Electrical Engineering and Computer Science, Texas A&M University–Kingsville, Kingsville, TX 78363, USA

Jun 10, 2026J/K move · Enter open · S save

MatchLM2Lite: A Scalable MLLM-to-Lite Framework for Reproduced Content Identification

Xiaotian Fan, Hiok Hian Ong, David Yuchen Wang, Zirui Zhu+2

Tiktok Singapore, Singapore · Tiktok National University of Singapore School of Computing Singapore, Singapore · Tiktok Content Intelligence San Jose, CA, USA+1

Multimodal Large Language Models

Open paper AskPDF arXiv

Content moderation is critical for online video platforms to ensure content safety, protect creators, and sustain positive user experiences. Beyond filtering harmful content, platforms must guarantee content authenticity at scale so that users are exposed to diverse, original videos rather than low-value reproductions. We present MatchLM2Lite, a real-time, production-grade reproduced content identification (RCI) system that leverages the powerful understanding of a multimodal large language model (MLLM) distilled into a small and fast-inference model. Our system jointly models video, audio, and text signals, operating on pairs of videos to produce fine-grained reproduction scores. The system comprises two modules, MatchLM and MatchLite, and a two-stage training recipe. First, our high-capacity MLLM, MatchLM, serves as a teacher model to define the upper bound of RCI performance. Its capabilities are then distilled into a compact student model, MatchLite. This design allows MatchLite to deliver low-latency, high-throughput inference on video pairs while preserving much of MatchLM's accuracy, making it suitable for integration into real-time recommendation systems. MatchLM achieves an F1-score improvement of +8.57 compared to our previous production model. After knowledge distillation, MatchLite retains a +6.55 gain in F1-score while reducing computational cost by 35x. Deployed at scale, MatchLM2Lite enables efficient, pairwise multimodal RCI, stably serving online traffic at high queries per second (QPS) with an end-to-end latency below 30 seconds. This system has reduced the reproduced video view rate on our platform by 2.5% without degrading user engagement, demonstrating its effectiveness in a large-scale production environment.

Abstract

Video Representation Learning

Explore similar work

MatchLM2Lite: A Scalable MLLM-to-Lite Framework for Reproduced Content Identification

MLT-Dedup: Efficient Large-Scale Online Video Deduplication via Multi-Level Representations and Spatial-Temporal Matching

Copyright Leakage Detection in AI Generated Videos