facebookresearch / instpred

Predicting Future Instance Segmentation by Forecasting Convolutional Features
Other
79 stars 32 forks source link

Doubt about auto-regressive fine tuning for short term prediction #7

Open guptasonam1602 opened 4 years ago

guptasonam1602 commented 4 years ago

Hi, As per the paper for short term prediction, you are predicting only a single frame subsampled by a factor of 3 (i.e. for input as {X(t-9), X(t-6), X(t-3, X(t)} you will predict the frame at X_(t+3). In table 2 different values of the AP50 metric are for F2F w/o ar. fine-tuning and with ar. fine-tuning for short term prediction. But should the values not be the same? For predicting a single frame only, how does autoregressive criterion make any difference?