Synergized 3D object and lane/road detection with association and temporal aggregation using graph neural networks

1 . An apparatus for image processing, comprising:

at least one memory; and

at least one processor coupled to the at least one memory, the at least one processor, individually or in any combination, is configured to:

detect (1) a set of polylines from at least one of a set of bird's eye view (BEV) features or a set of perspective view (PV) features associated with a set of images and (2) a set of three-dimensional (3D) objects in the set of BEV features;

associate the set of polylines with the set of 3D objects;

construct at least one graph representing the set of polylines and the set of 3D objects by creating a plurality of nodes and a plurality of edges to represent the set of polylines and the set of 3D objects;

aggregate the constructed at least one graph with other sets of graphs created from other sets of images at different times to capture a long-term dependency between the set of polylines and the set of 3D objects;

update the set of polylines based on a set of nearby 3D objects or update the set of 3D objects based on a set of nearby polylines; and

output an indication of the updated set of polylines or the updated set of 3D objects.

2 . The apparatus of claim 1 , wherein the at least one processor, individually or in any combination, is further configured to:

extract the set of PV features from the set of images prior to the detection; and

convert the set of PV features to the set of BEV features.

3 . The apparatus of claim 1 , wherein to detect the set of polylines from the set of BEV features, the at least one processor, individually or in any combination, is configured to:

detect a set of lane boundaries from the set of BEV features using a map decoder; and

obtain the set of polylines based on the detected set of lane boundaries.

4 . The apparatus of claim 1 , wherein to detect (1) the set of polylines from at least one of the set of BEV features or the set of PV features associated with the set of images and (2) the set of 3D objects in the set of BEV features, the at least one processor, individually or in any combination, is configured to:

detect (1) the set of polylines from at least one of the set of BEV features or the set of PV features associated with the set of images and (2) the set of 3D objects in the set of BEV features using at least one convolutional neural network (CNN).

5 . The apparatus of claim 1 , wherein each node in the plurality of nodes includes property information of a polyline in the set of polylines or a 3D object in the set of 3D objects.

6 . The apparatus of claim 5 , wherein the property information includes at least one of: a location, a size, or an orientation of the polyline or the 3D object.

7 . The apparatus of claim 1 , wherein to construct the at least one graph representing the set of polylines and the set of 3D objects by creating the plurality of nodes and the plurality of edges to represent the set of polylines and the set of 3D objects, the at least one processor, individually or in any combination, is configured to:

compute a set of distances between feature vectors of the plurality of nodes representing the set of polylines and the set of 3D objects; and

construct the plurality of edges based on the computed set of distances, wherein each edge in the plurality of edges connects to at least two nodes in the plurality of nodes.

8 . The apparatus of claim 1 , wherein to aggregate the constructed at least one graph with the other sets of graphs created from the other sets of images at the different times, the at least one processor, individually or in any combination, is configured to:

aggregate the constructed at least one graph with the other sets of graphs created from the other sets of images at the different times using at least one of a long short-term memory (LSTM) mechanism or an attention mechanism.

9 . The apparatus of claim 1 , wherein to update the set of polylines based on the set of nearby 3D object, the at least one processor, individually or in any combination, is configured to:

update the set of polylines based on the set of nearby 3D objects using a graph neural network (GNN) with an attention mechanism.

10 . The apparatus of claim 1 , wherein to update the set of 3D objects based on the set of nearby polylines, the at least one processor, individually or in any combination, is configured to:

update the set of 3D objects based on the set of nearby polylines using a graph neural network (GNN) with an attention mechanism.

11 . The apparatus of claim 1 , wherein to output the indication of the updated set of polylines or the updated set of 3D objects, the at least one processor, individually or in any combination, is configured to:

output, to a navigation application or an autonomous driving application, the indication of the updated set of polylines or the updated set of 3D objects.

12 . The apparatus of claim 1 , wherein to output the indication of the updated set of polylines or the updated set of 3D object, the at least one processor, individually or in any combination, is configured to:

store, in a memory or a cache, the indication of the updated set of polylines or the updated set of 3D objects.

13 . A method of image processing, comprising:

detecting (1) a set of polylines from at least one of a set of bird's eye view (BEV) features or a set of perspective view (PV) features associated with a set of images and (2) a set of three-dimensional (3D) objects in the set of BEV features;

associating the set of polylines with the set of 3D objects;

constructing at least one graph representing the set of polylines and the set of 3D objects by creating a plurality of nodes and a plurality of edges to represent the set of polylines and the set of 3D objects;

aggregating the constructed at least one graph with other sets of graphs created from other sets of images at different times to capture a long-term dependency between the set of polylines and the set of 3D objects;

updating the set of polylines based on a set of nearby 3D objects or updating the set of 3D objects based on a set of nearby polylines; and

outputting an indication of the updated set of polylines or the updated set of 3D objects.

14 . The method of claim 13 , further comprising:

extracting the set of PV features from the set of images prior to the detection; and

converting the set of PV features to the set of BEV features.

15 . The method of claim 13 , wherein detecting the set of polylines from the set of BEV features comprises:

detecting a set of lane boundaries from the set of BEV features using a map decoder; and

obtaining the set of polylines based on the detected set of lane boundaries.

16 . The method of claim 13 , wherein detecting (1) the set of polylines from at least one of the set of BEV features or the set of PV features associated with the set of images and (2) the set of 3D objects in the set of BEV features comprises:

detecting (1) the set of polylines from at least one of the set of BEV features or the set of PV features associated with the set of images and (2) the set of 3D objects in the set of BEV features using at least one convolutional neural network (CNN).

17 . The method of claim 13 , wherein each node in the plurality of nodes includes property information of a polyline in the set of polylines or a 3D object in the set of 3D objects.

18 . The method of claim 17 , wherein the property information includes at least one of: a location, a size, or an orientation of the polyline or the 3D object.

19 . The method of claim 13 , wherein constructing the at least one graph representing the set of polylines and the set of 3D objects by creating the plurality of nodes and the plurality of edges to represent the set of polylines and the set of 3D objects comprises:

computing a set of distances between feature vectors of the plurality of nodes representing the set of polylines and the set of 3D objects; and

constructing the plurality of edges based on the computed set of distances, wherein each edge in the plurality of edges connects to at least two nodes in the plurality of nodes.

20 . The method of claim 13 , wherein aggregating the constructed at least one graph with the other sets of graphs created from the other sets of images at the different times comprises:

aggregating the constructed at least one graph with the other sets of graphs created from the other sets of images at the different times using at least one of a long short-term memory (LSTM) mechanism or an attention mechanism.

21 . The method of claim 13 , wherein updating the set of polylines based on the set of nearby 3D objects comprises:

updating the set of polylines based on the set of nearby 3D objects using a graph neural network (GNN) with an attention mechanism.

22 . The method of claim 13 , wherein updating the set of 3D objects based on the set of nearby polylines comprises:

updating the set of 3D objects based on the set of nearby polylines using a graph neural network (GNN) with an attention mechanism.

23 . The method of claim 13 , wherein outputting the indication of the updated set of polylines or the updated set of 3D objects comprises:

outputting, to a navigation application or an autonomous driving application, the indication of the updated set of polylines or the updated set of 3D objects.

24 . The method of claim 13 , wherein outputting the indication of the updated set of polylines or the updated set of 3D objects comprises:

storing, in a memory or a cache, the indication of the updated set of polylines or the updated set of 3D objects.

25 . An apparatus for image processing, comprising:

means for detecting (1) a set of polylines from at least one of a set of bird's eye view (BEV) features or a set of perspective view (PV) features associated with a set of images and (2) a set of three-dimensional (3D) objects in the set of BEV features;

means for associating the set of polylines with the set of 3D objects;

means for constructing at least one graph representing the set of polylines and the set of 3D objects by creating a plurality of nodes and a plurality of edges to represent the set of polylines and the set of 3D objects;

means for aggregating the constructed at least one graph with other sets of graphs created from other sets of images at different times to capture a long-term dependency between the set of polylines and the set of 3D objects;

means for updating the set of polylines based on a set of nearby 3D objects or updating the set of 3D objects based on a set of nearby polylines; and

means for outputting an indication of the updated set of polylines or the updated set of 3D objects.

26 . A non-transitory computer-readable medium storing computer executable code, the code when executed by at least one processor causes the at least one processor to:

detect (1) a set of polylines from at least one of a set of bird's eye view (BEV) features or a set of perspective view (PV) features associated with a set of images and (2) a set of three-dimensional (3D) objects in the set of BEV features;

associate the set of polylines with the set of 3D objects;

construct at least one graph representing the set of polylines and the set of 3D objects by creating a plurality of nodes and a plurality of edges to represent the set of polylines and the set of 3D objects;

aggregate the constructed at least one graph with other sets of graphs created from other sets of images at different times to capture a long-term dependency between the set of polylines and the set of 3D objects;

update the set of polylines based on a set of nearby 3D objects or update the set of 3D objects based on a set of nearby polylines; and

output an indication of the updated set of polylines or the updated set of 3D objects.