Hello,
Unlike RNN and LSTM transfer the input of time series to a "state" when evaluation, TCN still takes in "time series". Therefore, TCN needs more memory when evaluation.
And also unlike forecasting, classification doesn't need to output time series.
Even though the performance might be better than the former ones, I can't explain why I use TCN.
It seems states of LSTM might be more explainable.
Could anyone help me to figure out the reason?
The working of TCN is based on the fact that if the proposal bounding boxes adjacent to the current frame have high detection scores, it is very likely to have the high probability for the presence of same object in the current frame located at a relative position with the other frames.
TCN is trained with the features of region proposals optimized through NMS and max-pooling. The TCN takes time series inputs containing detection scores, tracking scores and anchor offsets and generates a list probabilities as output which indicates whether the proposal contains the object of defined classes. With the temporal features, the TCN generates the prediction of classification in subsequent frames. The re-scoring is done by incorporating this information gives an impressive result even in the presence of temporal fluctuations. The network provides more flexibility for making changes in the size of receptive fields by increasing the number of convolutional layers or by introducing larger filter size.
Look in the work below if you can find any answer to your question.
https://www.researchgate.net/publication/333871785_Temporal_Convolutional_Networks_for_Anomaly_Detection_in_Time_Series
Article An Experimental Review on Deep Learning Architectures for Ti...
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series