A Large-scale TV Dataset for Partial Video Copy Detection

Abstract

This paper is interested with the performance evaluation of the partial video copy detection. Several public datasets exist designed from web videos. The detection problem is inherent to the continuous video broadcasting. The alternative is then to process with TV datasets offering a deeper scalability and a control of degradations for a fine performance evaluation. We propose in this paper a TV dataset called STVD. It is designed with a protocol ensuring a scalable capture and robust groundtruthing. STVD is the largest public dataset on the task with a near 83k videos having a total duration of 10, 660 hours. Performance evaluation results of representative methods on the dataset are reported in the paper for a baseline comparison.

Performance Characterization of 2D CNN Features for Partial Video Copy Detection

“…. , Xn} set of feature vectors • VGGNet [9,11,13,15,18,19,21,22] • ResNet [4,7,8,11,15] • InceptionNet [1,11,12,17] • AlexNet [1,10,12,21] Feature extraction • Fully connected layers [1,10,11,12,13,18,21] • Convolutional layers [1,4,7,8,11,12,15,17,19,22] • Low-dimensional [4,6,18,19] • RoI based features [8,22] Video matching…”

A large-scale TV video and metadata database for French political content analysis and fact-checking

“…In our work, we have used the TV workstation depicted in [20]. This workstation has been mainly used for the task of realtime video processing [20][21][22]. We detail here how we have adapted it for the video capture of political TV programs.…”

“…For storage optimization, we have set the video resolution to 320 × 240 for the capture with a parameter of 560 Kbps. This resolution has been set similar to [22,25] and presented as the best trade-off between the memory cost and video degradation. Suitable for most of the computer vision tasks such as the face and human activity recognition [28,29], it also fits with the workstation storage capacity of 5 TB and the campaign requirements of 3.4 TB.…”

“…These repeated programs can be detected and grouped automatically to support the selection. For the detection, we have used the method [22] that leverages NLP techniques and hashing to generate a unique hashcode for every repeated program.…”

“…One last problem remains, indeed TV suffers from latency [25] implying a shift in the time obtained in the metadata. To deal with this issue, we have used the approach of [22]. The French TV latency is modeled as a gaussian distribution such as L ∈ [L min , L max ] with L min < 0, L max > 0 and |L min | + L max ≃ 20 minutes.…”