Top-1 Solution for the 2019 CCF Big Data & Computing Intelligence Competition: Video Copyright Detection

The 2019 CCF Big Data & Computing Intelligence Competition (video copyright detection track) was jointly organized by iQIYI and CCF. The Hengyang Data team achieved the TOP1 result using a deep convolutional neural network (VGG16) to extract image feature vectors, performing approximate nearest‑neighbor search to find similar video frames, and then matching the time series of frames to pinpoint the start and end times of infringing segments with a precision of three seconds.

Core Task: Identify identical or visually similar video fragments (copy segments) between a query video and a reference video. Similarity includes various transformations such as cropping, saturation changes, subtitles, logos, and editing.

The competition provides three datasets: a reference set of 205 copyright videos (average 40 min each), and training and test sets consisting of 2–4 min clips containing a 1‑minute copy segment embedded at the beginning or middle.

The challenge lies in severe video transformations, especially extreme cropping where large portions of the original frame are removed.

Algorithm Overview:

1. Feature Extraction: Frames are decoded from videos, and a VGG16‑based convolutional feature combined with a Gaussian‑R‑MAC weighting emphasizes central image regions to handle cropping and quality degradation.

2. Search Algorithm: Due to the massive data volume, a graph‑based (NSG) nearest‑neighbor search is employed, achieving millisecond‑level single‑frame retrieval.

3. Video Matching: For each query frame, the top‑k nearest neighbors from the reference library are retrieved, and a matching algorithm aggregates these results to determine the copy segment.

The team observed that while the feature extraction works well for most transformations, extreme cropping and severe quality loss still degrade performance. By strengthening the weight of central image features using a 2‑D Gaussian distribution combined with RMAC, they improved the score by over ten points, reaching 95 points.

The video matching component acts like an expert system, compensating for weaknesses in earlier stages.

Insights & Reflections: The solution’s strengths are simplicity, portability, and commercial viability. Success was attributed to thorough preparation, deep understanding of the research direction, and robust hardware capable of handling tens of millions of decoded frames.

The team also highlighted the collaborative effort, long hours of debugging, and the rewarding experience of presenting their work.

Team Introduction: Hengyang Data Team, based in Shenzhen, consists of Li Chao (team lead), Chen Longsen (chairman), Liu Zhou, and Huang Yuxiang. Their expertise spans high‑performance large‑scale video/image retrieval, video copyright detection, illegal video blocking, image‑to‑video search, and person re‑identification.