Canonical Correlation-Based Feature Fusion Approach for Scene Classification

Loss Switching Fusion with Similarity Search for Video Classification

“…This improvement is gained from the loss switching mechanism which forces the learned representation to be more discrimi- native and compact. It should also be noted that LSFNet performs better than CCA [3], which is the feature fusion technique using correlation analysis. With LSH, the classification performance increases by 4%.…”

“…We also compare the performance of our fusion with PCA and autoencoder alone by truncating the features down to the same dimension (128D) as LSFNet. For PCA, this leads to a loss of about 9% feature 88.9 90.7 † Our own pipeline using Fisher score for each spatiotemporal descriptor followed by Canonical Correlation Analysis (CCA) [3] for the feature fusion. ‡ Our own pipeline using DT [26] followed by Fisher vector (FV) [37,38], then Fisher score is used to select the top-50% feature components for LSH.…”

“…Secondly, some feature extraction techniques [7][8][9] ignore the relationship between spatial and temporal information as they extract such information separately, which leads to the loss of information. Thirdly, scene understanding needs to fuse both background and foreground motions to represent the information within a given video [3,10], and this is a challenging research area that has not been fully explored. * The author performed the work while he was a Computer Vision Research Intern at iCetana Pty Ltd.…”

“…Scene understanding [1][2][3][4][5][6] in noisy videos is a challenging task. Firstly, videos contain too much redundant information, which slows down the training process and makes the feature extraction much harder.…”