To reduce the number of comparisons, we construct an inverted index to identify the videos which have a common visual word. Video similarity is inferred by calculating cosine similarity using tf-idf (term frequency / inverse-document frequency) between the two video histograms. Correspondingly, a video level histogram is generated by summing oveer the individual keyframe histograms. K = 1000 gives best results from experiments.įor each keyframe in a video, the Nearest cluster is identified to generate a keyframe level histogram. A sample of random 100K frames are used for visual codebook generation. Online Mini-batch K means is used to generate the codebook clusters. The video-level feature vector is calculated by concatenating over the individual keyframe feature vectors.Ī visual codebook is generated using the above feature vectors. Each frame is then represented by a feature vector of 1376 dimensions. Max-pooling is applied on the intermediate feature maps to extract one single value. Each video keyframe is forward passed through the intermediate layers of AlexNet to get frame feature vector. Here, pretrained weights of AlexNet is used to extract feature vectors. Pretrained CNN networks have been proven to work well on many vision tasks such as Classification, Segmentation etc. This is done in following phasesĬonvolution Neural Networks are used to extract features from the video keyframes. The problem is tackled by using a bag-of visual words model for each video. The dataset already provides 398,008 keyframes extracted from the videos by shot boundary detection method. The dataset contains a total of 13,129 videos and 24 query videos. The standard for NDVD tasks is the CC_WEB_VIDEO dataset. Given a corpus of videos, identify if a query video is a near duplicate of an existing video in corpus. Near duplicates videos are videos which have same visual content but differ in format ( scale, transformation, encoding etc ) or have small content modifications ( color / lightning / small text superimposition etc). Near duplicate videos are a bigger class of problems which cover duplicate videos. ![]() It has become essential that we build robust systems to detect and purge these duplicates. According to a recent study, 31.7% of videos on Youtube are duplicates with duplicates occupying 24% of storage space. A Cisco forecast estimates that videos will constitute 80% of internet traffic by 2019. Currently, Youtube reports 500h of video content being uploaded every minute. With the explosion of social networks, video content has risen exponentially over the last few years.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |