Joint forecasting of feature and feature motion

1. A computer-implemented method of forecasting the semantic output of at least one frame, the method comprising:

receiving a plurality of input frames from a camera up to a predetermined time;

processing via a down-sampling module of a neural network the plurality of input frames to receive a plurality of feature tensors;

determining spatio-temporal correlations between the plurality of feature tensors;

processing the plurality of feature tensors and the spatio-temporal correlations to receive at least one forecasted feature; and

processing via an up-sampling module of the neural network the at least one forecasted feature tensor to receive at least one forecasted semantic output for a time larger than the predetermined time;

wherein in the processing of the plurality of feature tensors and the spatio-temporal correlations to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by direct regression based on the plurality of feature tensors and the spatio-temporal correlations.

2. The method according to claim 1 , further comprising:

anticipating at least one anticipated future event occurring after the predetermined time, based on the at least one forecasted semantic output; and

making a decision based on the at least one anticipated future event.

3. The method according to claim 1 , wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by warping each of the plurality of feature tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurality of forecasted counterparts into the at least one forecasted feature tensor using predetermined weight vectors.

4. The method according to claim 1 , wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by:

performing direct regression based on the plurality of feature tensors and the spatio-temporal correlations to receive at least one first auxiliary forecasted feature tensor;

warping each of the plurality of feature tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurality of forecasted counterparts into at least one second auxiliary forecasted feature using predetermined weight vectors;

inferring tensor blending weights from the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor; and

blending the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor using the tensor blending weights into the at least one forecasted feature tensor.

5. An apparatus configured for executing the method of claim 1 , the apparatus comprising:

the camera;

the down-sampling module;

the up-sampling module; and

a processor.

6. The apparatus according to claim 5 , wherein the method further comprises:

anticipating at least one anticipated future event occurring after the predetermined time, based on the at least one forecasted semantic output; and

making a decision based on the at least one anticipated future event.

7. The apparatus according to claim 5 , wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by warping each of the plurality of feature tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurlaity of forecasted counterparts into the at least one forecasted feature tensor using predetermined weight vectors.

8. The apparatus according to claim 5 , wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature tensor, the at least one forecasted feature tensors is received by:

performing direct regression based on the plurality of features tensors and the spatio-temporal correlations to receive at least one first auxiliary forecasted feature tensor;

warping each of the plurality of features tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurality of forecasted counterparts into at least one second auxiliary forecasted feature using predetermined weight vectors;

inferring tensor blending weights from the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor; and

blending the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor using the tensor blending weights into the at least one forecasted feature tensor.

9. A vehicle component, comprising a processor and a memory having access to instruction that, when provided to the processor, causes the processor to execute the method of claim 1 .

10. The vehicle component according to claim 9 , wherein the method further comprises:

anticipating at least one anticipated future event occurring after the predetermined time, based on the at least one forecasted semantic output; and

making a decision based on the at least one anticipated future event.

11. The vehicle component according to claim 9 ,

wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by warping each of the plurality of feature tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurality of forecasted counterparts into the at least one forecasted feature tensor using predetermined weight vectors.

12. The vehicle component according to claim 9 , wherein in the processing of the plurality of feature tensors to receive the at least one forecasted feature, the at least on forecasted feature tensors is received by:

performing direct regression based on the plurality of feature tensors and the spatio-temporal correlations to receive at least one first auxiliary forecasted feature tensor;

warping each of the plurality of feature tensors into a forecasted counterpart to obtain a plurality of forecasted counterparts, and then blending the plurality of forecasted counterparts into at least one second auxiliary forecasted feature using predetermined weight vectors;

inferring tensor blending weights from the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor; and

blending the at least one first auxiliary forecasted feature tensor and the at least one second auxiliary forecasted feature tensor using the tensor blending weights into the at least one forecasted feature tensor.

13. A vehicle component, comprising a processor and a memory having access to instruction that when provided to the processor, causes the processor to simultaneously execute a method for single-frame prediction and a method for dense semantic forecasting of at least one future frame, wherein:

the method for single-frame prediction comprises:

receiving the input frame from a camera up to a predetermined time,

processing the input frame via a down-sampling module of a neural network to receive a corresponding feature tensor, and

caching the feature tensors for later use, and processing the corresponding feature tensor via an up-sampling module of the neural network to receive the semantic output; and

the method for dense semantic forecasting comprises of:

retrieving a plurality of cached feature tensors up to the predetermined time,

determining spatio-temporal correlations between the plurality of feature tensors,

processing the plurality of cached feature tensors and the spatio-temporal correlations to receive at least one forecasted feature tensor, and

processing via an up-sampling module of the neural network the at least one forecasted feature tensor to receive at least one forecasted semantic output for a time larger than the predetermined time;

wherein the processing of the plurality of cached feature tensors and the spatio-temporal correlations to receive the at least one forecasted feature tensor, the at least one forecasted feature tensor is received by direct regression based on the plurality of cached feature tensors and the spatio-temporal correlations.