The main purpose Keras TimeDistributed layer is fully connected at the time dimension.
For example Faster RCNN, 1 Photo 16 generates ROI, the need for classification and regression for each ROI, the ROI is the dimension 7 × 7 × 512, the length and width is 7,512 channels, dimension 16 is the ROI 16 × 7 × 7 × 512, 16 require classification and regression, this time can be used TimeDistributed layer of disposable input 16 × 7 × 7 × 512, full connection, corresponding to 16 time slices, each 7 7 × 512 × sections are fully connected, respectively (Dense), 16 finally obtain a classification and regression.